Анализ сайта researchforum.microsoft.com
Основное Готовность: 30%
Домен
researchforum.microsoft.com
Состояние доменного имени
?
Проверяем корректность доменного имени и наличие технических проблем на уровне домена.
Используйте для продвижения только домен второго уровня.
Длина домена велика. Но если вы продвигаете запрос, входящий в название домена, то это хорошо.
Ответ сервера
200 Успешный ответ
HTTP-код ответа и цепочка редиректов
?
Код 200 — страница доступна. Коды 3xx — редиректы (цепочки замедляют загрузку и размывают ссылочный вес). Коды 4xx/5xx — ошибки, поисковик не сможет проиндексировать страницу.
Кол-во редиректов 3 слишком большое! Проверьте настройки сайта и веб-сервера!
Цепочка редиректов:
http://researchforum.microsoft.com
301 MovedPermanently
https://www.microsoft.com/en-us/research/event/microsoft-research-forum
301 MovedPermanently
https://www.microsoft.com/en-us/research/event/microsoft-research-forum/
200 OK
Безопасность
Сайт безопасен
Использование HTTPS и SSL-сертификат
?
HTTPS — обязательный стандарт. Google и Яндекс отдают предпочтение защищённым сайтам. Отсутствие SSL или просроченный сертификат ведут к предупреждениям в браузере и снижению позиций.
На сайте работает защищенный протокол ssl и сайт открывается по https.
Ssl-сертификат действителен до 09.01.2027 2:59:59.
Включён HSTS (Strict-Transport-Security) — защита от подмены на http.
HTTP автоматически перенаправляется на HTTPS.
Поздравляем! Сайт не содержится в реестре РКН.
Кодировка
Кодировка символов страницы
?
Стандарт — UTF-8. Неправильная кодировка вызывает нечитаемые символы и мешает поисковику корректно распознать текст страницы.
Кодировка на странице не определена. Укажите явно кодировку документа!
Язык
en-US
Атрибут lang в HTML-теге
?
Атрибут lang (<html lang="ru">) сообщает поисковикам и браузерам, на каком языке написана страница. Помогает при ранжировании в региональном поиске.
Язык документа указан явно: en-US.
Скорость загрузки
~1,73сек
Время отклика сервера (TTFB)
?
Time To First Byte — время до получения первого байта от сервера. Норма до 200 мс. Медленный отклик ухудшает пользовательский опыт и ранжирование: Яндекс и Google учитывают скорость страниц.
Скорость загрузки сайта 1,73сек превышает 1 секунду. Желательно улучшить работу сайта!
Объем документа
570Кб
Размер HTML-кода страницы
?
Слишком большой HTML замедляет парсинг браузером и сканирование поисковым роботом. Рекомендуется не более 200 Кб.
Объем html-документа 570Кб оптимален.
Структура html-документа корректна.
Ресурсы
Ресурсы: 43
Внешние ресурсы страницы (CSS, JS, изображения)
?
Количество и тип подключённых ресурсов влияют на скорость загрузки. Большое число запросов увеличивает время рендеринга страницы.
Кол-во файлов ресурсов 43 много для одной страницы. Приемлемо до 10. Проведите оптимизацию файлов ресурсов!
Показать полный список ресурсов
| Тип | Название | Значение |
|---|---|---|
| stylesheet | https://www.microsoft.com/en-us/research/wp-content/plugins/column-shortcodes/assets/css/shortcodes.css?ver=1.0.1 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-content/plugins/moray-blocks/dist/css/shared-style.css?ver=0.2.0 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-content/plugins/moray-blocks/dist/css/style.css?ver=0.2.0 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-content/plugins/msr-blocks-library/dist/css/shared.css?ver=1787066359 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-content/plugins/msr-blocks-library/dist/css/frontend.css?ver=1787066359 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-content/plugins/taxonomy-images/css/style.css?ver=0.9.6 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-content/plugins/elasticpress/dist/css/general-styles.css?ver=66295efe92a630617c00 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/css/microsoft-research-moray.min.css?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-includes/css/dist/components/style.min.css?ver=7.0.4 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-includes/css/dist/preferences/style.min.css?ver=7.0.4 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-includes/css/dist/block-editor/style.min.css?ver=7.0.4 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-includes/css/dist/reusable-blocks/style.min.css?ver=7.0.4 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-includes/css/dist/patterns/style.min.css?ver=7.0.4 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-includes/css/dist/editor/style.min.css?ver=7.0.4 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/css/blocks-style.min.css?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-content/plugins/elasticpress/dist/css/autosuggest-styles.css?ver=d87f34a78edccbda21b1 | |
| stylesheet | https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/vendor/duet-date-picker/duet/themes/default.css?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332 | |
| stylesheet | https://uhf.microsoft.com/statics/20260814.11.18/css/style-By05NU7M.css | |
| js | https://www.microsoft.com/en-us/research/wp-includes/js/jquery/jquery.min.js?ver=3.7.1 | |
| js | https://www.microsoft.com/en-us/research/wp-includes/js/jquery/jquery-migrate.min.js?ver=3.4.1 | |
| js | https://js.monitor.azure.com/scripts/c/ms.analytics-web-3.min.js | |
| js | module | https://uhf.microsoft.com/statics/20260814.11.18/js/entry.js |
| js | https://wcpstatic.microsoft.com/mscc/lib/v2/wcp-consent.js | |
| js | https://www.microsoft.com/en-us/research/wp-content/plugins/moray-blocks/dist/js/shared.js?ver=0.2.0 | |
| js | https://www.microsoft.com/en-us/research/wp-content/plugins/moray-blocks/dist/js/frontend.js?ver=0.2.0 | |
| js | https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/mwf/bundle.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332 | |
| js | https://www.microsoft.com/en-us/research/wp-content/plugins/msr-blocks-library/dist/js/shared.js?ver=0.3.0 | |
| js | https://www.microsoft.com/en-us/research/wp-content/plugins/msr-blocks-library/dist/js/frontend.js?ver=ed590bdf264223835ab6 | |
| js | https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/accessible-tabs.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332 | |
| js | https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/clamp.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332 | |
| js | https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/responsive-tables.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332 | |
| js | https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/wedecs.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332 | |
| js | https://www.microsoft.com/en-us/research/wp-includes/js/dist/dom-ready.min.js?ver=a06281ae5cf5500e9317 | |
| js | https://www.microsoft.com/en-us/research/wp-includes/js/dist/hooks.min.js?ver=7496969728ca0f95732d | |
| js | https://www.microsoft.com/en-us/research/wp-includes/js/dist/i18n.min.js?ver=781d11515ad3d91786ec | |
| js | https://www.microsoft.com/en-us/research/wp-includes/js/dist/a11y.min.js?ver=af934e5259bc51b8718e | |
| js | https://www.microsoft.com/en-us/research/wp-includes/js/dist/url.min.js?ver=bb0f766c3d2efe497871 | |
| js | https://www.microsoft.com/en-us/research/wp-includes/js/dist/api-fetch.min.js?ver=d7efe4dc1468d36c39b8 | |
| js | https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/microsoft-research.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332 | |
| js | https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/vendor/duet-date-picker/duet/duet.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332 | |
| js | https://www.microsoft.com/en-us/research/wp-content/plugins/microsoft-metrics/assets/js/consent-manager.js?ver=1.3.0 | |
| js | https://www.microsoft.com/en-us/research/wp-content/plugins/microsoft-uhf/assets/microsoft-uhf.js?ver=0.6.1 | |
| js | https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/search.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332 |
Серверные заголовки
Кол-во: 15
HTTP-заголовки ответа сервера
?
Заголовки сервера передают браузеру и поисковику служебную информацию: кеширование, безопасность (CSP, HSTS), сжатие (gzip). Правильная настройка ускоряет загрузку и повышает защищённость.
Найдены серверные заголовки 15шт. Подробнее про серверные заголовки.
Показать полный список серверных заголовков
| Ключ | Значение |
|---|---|
| X-Frame-Options | SAMEORIGIN |
| Link | <https://www.microsoft.com/en-us/research/wp-json/>; rel="https://api.w.org/";<https://www.microsoft.com/en-us/research/wp-json/wp/v2/msr-event/1140163>; rel="alternate"; title="JSON"; type="application/json";<https://www.microsoft.com/en-us/research/?p=1140163>; rel=shortlink |
| X-ElasticPress-Query | true |
| x-azure-ref | 20260822T071447Z-1659b98fb99n485lhC1FRA7n500000000u00000000007aen |
| X-AFD | WWWRule |
| Date | Sat, 22 Aug 2026 07:14:48 GMT |
| Transfer-Encoding | chunked |
| Connection | keep-alive |
| Set-Cookie | bStore=Y; expires=Sat, 22-Aug-2026 07:14:58 GMT |
| TLS_version | tls1.2 |
| Strict-Transport-Security | max-age=31536000; includeSubDomains |
| ms-cv | CASMicrosoftCVbf008b44.0 |
| ms-cv-esi | CASMicrosoftCVbf008b44.0 |
| X-EdgeScape-Location | country_code=RU |
| X-RTag | OneRoute_Default |
CMS
WordPress
Система управления сайтом (движок)
?
CMS — это движок, на котором работает сайт (WordPress, 1C-Bitrix, Tilda и др.). Знание CMS помогает понять возможности SEO-оптимизации и подобрать подходящие инструменты. «Не определена» — вероятно, самописный сайт или нестандартная сборка.
Сайт работает на CMS WordPress.
Веб-сервер
Не определён
Программное обеспечение сервера
?
Веб-сервер — это ПО, которое отдаёт страницы посетителям (nginx, Apache, IIS, LiteSpeed и др.). Определяется по серверным заголовкам ответа (Server, X-Powered-By и т.п.). «Не определён» — сервер намеренно скрывает эти заголовки, это нормальная практика безопасности.
Веб-сервер не определён — заголовок Server скрыт. Это не ошибка: часто так настраивают из соображений безопасности.
Мета-теги Готовность: 23%
Title
Microsoft Research Forum - Microsoft Research
Заголовок страницы в браузере и поисковой выдаче
?
Title — главный SEO-заголовок страницы. Влияет на CTR в поиске и ранжирование. Оптимальная длина: 50–70 символов. Ключевые слова — ближе к началу.
Устраните дубли в title: microsoft(2), research(2)
Число символов в title 45 оптимально (норма: от 40 до 45).
Description
This series explores recent research advances, bold new ideas, and important discussions with the global research community.
Описание страницы в поисковой выдаче (сниппет)
?
Meta Description — текст под заголовком в выдаче. Напрямую на позиции не влияет, но влияет на CTR. Оптимальная длина: 120–160 символов.
Число символов в description 124 оптимально (норма: от 120 до 130).
Keywords
Список ключевых слов страницы (устаревший тег)
?
Meta Keywords не учитывается Яндексом и Google для ранжирования с 2009–2012 годов. Заполнение не обязательно, но не вредит. Конкурент может использовать содержимое для анализа.
Установите мета-тег keywords!
Канонический Url
https://www.microsoft.com/en-us/research/event/microsoft-research-forum/
Указывает поисковику основную версию страницы
?
Canonical (rel=canonical) предотвращает проблему дублей страниц. Должен точно совпадать с URL проверяемой страницы. Неправильный canonical может передать ссылочный вес на другую страницу.
Канонический Url прописан корректно.
Robots
index, follow, max-image-preview:large, max-snippet:-1, max-video-preview:-1
Директивы для поисковых роботов на уровне страницы
?
Meta Robots управляет индексацией конкретной страницы: index/noindex — индексировать ли, follow/nofollow — следовать ли по ссылкам. Noindex полностью исключает страницу из поиска.
Meta-тег robots со значениями index, follow не накладывает ограничений на индексирование и показ контента.
Адаптивность
width=device-width, initial-scale=1
Настройка масштабирования на мобильных устройствах
?
Тег viewport (<meta name="viewport">) сообщает браузеру, как масштабировать страницу на мобильных. Стандарт: width=device-width, initial-scale=1. Отсутствие — признак отсутствия мобильной версии.
Meta-тег viewport со значением-константой width=device-width задаёт ширину страницы в соответствии с размером экрана.
Meta-тег viewport со значением initial-scale=1.0 определяет масштаб 1:1, т.е. «не масштабировать».
Разметка OpenGraph
Кол-во: 10
Мета-теги для красивых превью в соцсетях
?
OpenGraph (og:title, og:description, og:image) управляет тем, как страница выглядит при репосте в социальных сетях и мессенджерах. Отсутствие OG-тегов — невзрачный превью при шеринге.
Разметка OpenGraph задана. Страница оптимизирована под социальные сети.
Показать полный список og мета-тегов
| Тип | Значение |
|---|---|
| og:locale | en_US |
| og:type | article |
| og:title | Microsoft Research Forum - Microsoft Research |
| og:description | This series explores recent research advances, bold new ideas, and important discussions with the global research community. |
| og:url | https://www.microsoft.com/en-us/research/event/microsoft-research-forum/ |
| og:site_name | Microsoft Research |
| og:image | https://www.microsoft.com/en-us/research/wp-content/uploads/2025/05/Research-Forum-hero_1400x788.jpg |
| og:image:width | 1400 |
| og:image:height | 788 |
| og:image:type | image/jpeg |
Все мета-теги
Кол-во: 30
Полный список мета-тегов страницы
?
Таблица всех meta-тегов, включая нестандартные. Позволяет найти опечатки, дубли и лишние теги.
Найдены мета-теги 30шт. Мета-теги не видимы для человека и предназначены для обмена информацией между веб-страницей и поисковыми системами, браузерами и другими веб-службами. С ними роботы 🤖 и устройства ведут себя более ожидаемо.
Показать полный список мета-тегов
| Тип | Название | Значение |
|---|---|---|
| name | viewport | width=device-width, initial-scale=1 |
| name | twitter:dnt | on |
| name | awa-product | MSR |
| name | awa-stv | 9.7.0 |
| name | awa-sitesection | |
| name | awa-pageType | Event |
| name | awa-market | en-us |
| name | awa-env | Production |
| name | awa‐asst | 1140163 |
| name | awa-pgidx | 1 |
| name | awa-pgtot | -1 |
| name | awa-pgtop | Artificial intelligence; Medical, health and genomics |
| name | robots | index, follow, max-image-preview:large, max-snippet:-1, max-video-preview:-1 |
| name | description | This series explores recent research advances, bold new ideas, and important discussions with the global research community. |
| name | twitter:card | summary_large_image |
| name | twitter:site | @MSFTResearch |
| name | generator | WordPress 7.0.4 |
| name | research-area | Artificial intelligence; Medical, health and genomics |
| property | og:locale | en_US |
| property | og:type | article |
| property | og:title | Microsoft Research Forum - Microsoft Research |
| property | og:description | This series explores recent research advances, bold new ideas, and important discussions with the global research community. |
| property | og:url | https://www.microsoft.com/en-us/research/event/microsoft-research-forum/ |
| property | og:site_name | Microsoft Research |
| property | article:publisher | https://www.facebook.com/microsoftresearch/ |
| property | article:modified_time | 2026-08-06T17:09:05+00:00 |
| property | og:image | https://www.microsoft.com/en-us/research/wp-content/uploads/2025/05/Research-Forum-hero_1400x788.jpg |
| property | og:image:width | 1400 |
| property | og:image:height | 788 |
| property | og:image:type | image/jpeg |
Оптимизация Готовность: 70%
Структура
Ошибок нет
Семантические HTML-элементы страницы
?
Проверяет наличие основных структурных элементов: nav, header, footer, main. Корректная семантическая структура помогает поисковику понять архитектуру страницы.
Структура документа корректна (теги <html> и <body> присутствуют в единственном экземпляре).
Контент
Есть ошибки
Объём и качество текстового содержимого
?
Анализирует объём полезного текста на странице. Слишком мало — страница может считаться малополезной. Слишком много — ухудшается читаемость и восприятие.
Абзацев с текстом 9 слишком мало. Добавьте больше абзацев с текстом (тег <p>)!
Среднее число слов в абзаце 9 слишком мало. Сделайте контент более читаемым!
Кол-во слов 559 не очень много. Добавьте побольше текста (хотя бы 800 слов)!
Слова из title 3 встречаются в тексте достаточно.
Кол-во знаков контента 4141 на странице оптимально.
Заголовки
Ошибок нет
Иерархия заголовков H1–H6
?
H1 должен быть один и содержать ключевой запрос. H2–H6 описывают подразделы. Пропуск уровней (H1 → H3) и несколько H1 — типичные ошибки, снижающие понятность страницы для поисковика.
На странице присутствуют заголовки <h1> 1. Это прекрасно.
На странице присутствуют заголовки <h2> 3. Это хорошо.
На странице присутствуют заголовки <h3> 9.
Тошнота
6,63
Насколько одно слово доминирует в тексте
?
Классическая тошнота = √(частота самого повторяющегося слова). Норма до 7–8: текст воспринимается естественно. Выше — поисковик может счесть страницу переспамленной.
Тошнота превышает норму 3. Измените текст страницы!
Академич. тошнота
22,54%
Насколько текст перенасыщен ключевыми словами
?
Академическая тошнота = (частота слова / общее количество слов) × 100%. Показывает долю конкретного слова в тексте. Норма 5–15%.
Академическая тошнота превышает норму 5-15%. Измените текст страницы!
Семантическое ядро
20
Наиболее часто встречающиеся слова на странице
?
Топ слов по частоте использования. Показывает, какие слова доминируют в тексте с точки зрения поисковика.
Контент страницы содержит осмысленный текст и слова.
Показать список слов
| Слово | Кол-во | Частота |
|---|---|---|
| microsoft | 44 | 7,87% |
| research | 19 | 3,40% |
| privacy | 8 | 1,43% |
| education | 6 | 1,07% |
| explore | 5 | 0,89% |
| follow | 4 | 0,72% |
| surface | 4 | 0,72% |
| choices | 4 | 0,72% |
| software | 3 | 0,54% |
| security | 3 | 0,54% |
| health | 3 | 0,54% |
| careers | 3 | 0,54% |
| reality | 3 | 0,54% |
| partner | 3 | 0,54% |
| marketplace | 3 | 0,54% |
| series | 3 | 0,54% |
| copilot | 3 | 0,54% |
| people | 2 | 0,36% |
| human-computer | 2 | 0,36% |
| interaction | 2 | 0,36% |
Индексация Готовность: 0%
Индексирование
Ошибок нет
Разрешено ли индексирование страницы
?
Проверяет, не закрыта ли страница от индексации через robots.txt, meta robots или X-Robots-Tag. Страница, закрытая от индексации, не появится в поисковой выдаче.
Анкоров на странице 174 оптимально. Поисковые роботы обязательно проиндексируют сайт.
Robots.txt
Найден корректный robots.txt
Файл управления сканированием сайта роботами
?
Robots.txt указывает поисковым роботам, какие страницы сканировать, а какие — нет. Ошибки в файле могут случайно закрыть важные разделы от индексации.
Кол-во редиректов для файла robots.txt 3 слишком большое! Это может вызывать затруднение при индексации поисковыми роботами!
Проверяемая страница не запрещена в robots.txt.
Robots.txt доступен по постоянному адресу
Цепочка редиректов для файла robots.txt:
http://researchforum.microsoft.com/robots.txt
301 MovedPermanently
https://www.microsoft.com/en-us/research/event/microsoft-research-forum
301 MovedPermanently
https://www.microsoft.com/en-us/research/event/microsoft-research-forum/
200 OK
Показать содержимое robots.txt
<!DOCTYPE html>
<html lang="en-US" class="no-js">
<head>
<meta name="viewport" content="width=device-width, initial-scale=1" />
<meta name="twitter:dnt" content="on">
<script>document.documentElement.classList.remove('no-js');document.documentElement.classList.add('js');</script>
<script>
if ( -1 !== window.location.hash.indexOf( '#!' ) ) {
window.location.href = window.location.origin + window.location.pathname + window.location.hash.replace( '#!', '' ) + '/' + window.location.search;
}
if ( -1 !== window.location.hash.indexOf( '#' + 'past-episodes' ) ) {
window.location.href = window.location.origin + window.location.pathname + 'past-episodes' + '/' + window.location.search;
}
if ( -1 !== window.location.hash.indexOf( '#' + 'event-code-of-conduct' ) ) {
window.location.href = window.location.origin + window.location.pathname + 'event-code-of-conduct' + '/' + window.location.search;
}
if ( -1 !== window.location.hash.indexOf( '#' + 'faq' ) ) {
window.location.href = window.location.origin + window.location.pathname + 'faq' + '/' + window.location.search;
}
if ( -1 !== window.location.pathname.indexOf( '/overview' ) ) {
window.location.href = window.location.origin + window.location.pathname.replace( '/overview', '' ) + window.location.search;
}
</script>
<meta name="awa-product" content="MSR">
<meta name="awa-stv" content="9.7.0">
<meta name="awa-sitesection" content="">
<meta name="awa-pageType" content="Event">
<meta name="awa-market" content="en-us">
<meta name="awa-env" content="Production">
<meta name="awa‐asst" content="1140163">
<meta name="awa-pgidx" content="1">
<meta name="awa-pgtot" content="-1">
<meta name="awa-pgtop" content="Artificial intelligence; Medical, health and genomics">
<meta name='robots' content='index, follow, max-image-preview:large, max-snippet:-1, max-video-preview:-1' />
<!-- This site is optimized with the Yoast SEO plugin v27.8 - https://yoast.com/product/yoast-seo-wordpress/ -->
<title>Microsoft Research Forum - Microsoft Research</title>
<meta name="description" content="This series explores recent research advances, bold new ideas, and important discussions with the global research community." />
<link rel="canonical" href="https://www.microsoft.com/en-us/research/event/microsoft-research-forum/" />
<meta property="og:locale" content="en_US" />
<meta property="og:type" content="article" />
<meta property="og:title" content="Microsoft Research Forum - Microsoft Research" />
<meta property="og:description" content="This series explores recent research advances, bold new ideas, and important discussions with the global research community." />
<meta property="og:url" content="https://www.microsoft.com/en-us/research/event/microsoft-research-forum/" />
<meta property="og:site_name" content="Microsoft Research" />
<meta property="article:publisher" content="https://www.facebook.com/microsoftresearch/" />
<meta property="article:modified_time" content="2026-08-06T17:09:05+00:00" />
<meta property="og:image" content="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/05/Research-Forum-hero_1400x788.jpg" />
<meta property="og:image:width" content="1400" />
<meta property="og:image:height" content="788" />
<meta property="og:image:type" content="image/jpeg" />
<meta name="twitter:card" content="summary_large_image" />
<meta name="twitter:site" content="@MSFTResearch" />
<script type="application/ld+json" class="yoast-schema-graph">{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.microsoft.com\/en-us\/research\/event\/microsoft-research-forum\/","url":"https:\/\/www.microsoft.com\/en-us\/research\/event\/microsoft-research-forum\/","name":"Microsoft Research Forum - Microsoft Research","isPartOf":{"@id":"https:\/\/www.microsoft.com\/en-us\/research\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.microsoft.com\/en-us\/research\/event\/microsoft-research-forum\/#primaryimage"},"image":{"@id":"https:\/\/www.microsoft.com\/en-us\/research\/event\/microsoft-research-forum\/#primaryimage"},"thumbnailUrl":"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/05\/Research-Forum-hero_1400x788.jpg","datePublished":"2025-05-27T19:35:37+00:00","dateModified":"2026-08-06T17:09:05+00:00","description":"This series explores recent research advances, bold new ideas, and important discussions with the global research community.","breadcrumb":{"@id":"https:\/\/www.microsoft.com\/en-us\/research\/event\/microsoft-research-forum\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.microsoft.com\/en-us\/research\/event\/microsoft-research-forum\/"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.microsoft.com\/en-us\/research\/event\/microsoft-research-forum\/#primaryimage","url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/05\/Research-Forum-hero_1400x788.jpg","contentUrl":"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/05\/Research-Forum-hero_1400x788.jpg","width":1400,"height":788,"caption":"Research Forum | abstract background with colorful hexagons"},{"@type":"BreadcrumbList","@id":"https:\/\/www.microsoft.com\/en-us\/research\/event\/microsoft-research-forum\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.microsoft.com\/en-us\/research\/"},{"@type":"ListItem","position":2,"name":"Microsoft Research Forum"}]},{"@type":"WebSite","@id":"https:\/\/www.microsoft.com\/en-us\/research\/#website","url":"https:\/\/www.microsoft.com\/en-us\/research\/","name":"Microsoft Research","description":"","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.microsoft.com\/en-us\/research\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"}]}</script>
<!-- / Yoast SEO plugin. -->
<link rel='dns-prefetch' href='//www.microsoft.com' />
<link rel='dns-prefetch' href='//js.monitor.azure.com' />
<link rel='dns-prefetch' href='//wcpstatic.microsoft.com' />
<link rel='preconnect' href='https://wcpstatic.microsoft.com' />
<link rel="alternate" type="application/rss+xml" title="Microsoft Research » Feed" href="https://www.microsoft.com/en-us/research/feed/" />
<link rel="alternate" title="oEmbed (JSON)" type="application/json+oembed" href="https://www.microsoft.com/en-us/research/wp-json/oembed/1.0/embed?url=https%3A%2F%2Fwww.microsoft.com%2Fen-us%2Fresearch%2Fevent%2Fmicrosoft-research-forum%2F" />
<link rel="alternate" title="oEmbed (XML)" type="text/xml+oembed" href="https://www.microsoft.com/en-us/research/wp-json/oembed/1.0/embed?url=https%3A%2F%2Fwww.microsoft.com%2Fen-us%2Fresearch%2Fevent%2Fmicrosoft-research-forum%2F&format=xml" />
<style id="wp-img-auto-sizes-contain-inline-css">
img:is([sizes=auto i],[sizes^="auto," i]){contain-intrinsic-size:3000px 1500px}
/*# sourceURL=wp-img-auto-sizes-contain-inline-css */
</style>
<style id="wp-emoji-styles-inline-css">
img.wp-smiley, img.emoji {
display: inline !important;
border: none !important;
box-shadow: none !important;
height: 1em !important;
width: 1em !important;
margin: 0 0.07em !important;
vertical-align: -0.1em !important;
background: none !important;
padding: 0 !important;
}
/*# sourceURL=wp-emoji-styles-inline-css */
</style>
<style id="wp-block-library-inline-css">
:root{--wp-block-synced-color:#7a00df;--wp-block-synced-color--rgb:122,0,223;--wp-bound-block-color:var(--wp-block-synced-color);--wp-editor-canvas-background:#ddd;--wp-admin-theme-color:#007cba;--wp-admin-theme-color--rgb:0,124,186;--wp-admin-theme-color-darker-10:#006ba1;--wp-admin-theme-color-darker-10--rgb:0,107,160.5;--wp-admin-theme-color-darker-20:#005a87;--wp-admin-theme-color-darker-20--rgb:0,90,135;--wp-admin-border-width-focus:2px}@media (min-resolution:192dpi){:root{--wp-admin-border-width-focus:1.5px}}.wp-element-button{cursor:pointer}:root .has-very-light-gray-background-color{background-color:#eee}:root .has-very-dark-gray-background-color{background-color:#313131}:root .has-very-light-gray-color{color:#eee}:root .has-very-dark-gray-color{color:#313131}:root .has-vivid-green-cyan-to-vivid-cyan-blue-gradient-background{background:linear-gradient(135deg,#00d084,#0693e3)}:root .has-purple-crush-gradient-background{background:linear-gradient(135deg,#34e2e4,#4721fb 50%,#ab1dfe)}:root .has-hazy-dawn-gradient-background{background:linear-gradient(135deg,#faaca8,#dad0ec)}:root .has-subdued-olive-gradient-background{background:linear-gradient(135deg,#fafae1,#67a671)}:root .has-atomic-cream-gradient-background{background:linear-gradient(135deg,#fdd79a,#004a59)}:root .has-nightshade-gradient-background{background:linear-gradient(135deg,#330968,#31cdcf)}:root .has-midnight-gradient-background{background:linear-gradient(135deg,#020381,#2874fc)}:root{--wp--preset--font-size--normal:16px;--wp--preset--font-size--huge:42px}.has-regular-font-size{font-size:1em}.has-larger-font-size{font-size:2.625em}.has-normal-font-size{font-size:var(--wp--preset--font-size--normal)}.has-huge-font-size{font-size:var(--wp--preset--font-size--huge)}:root .has-text-align-center{text-align:center}:root .has-text-align-left{text-align:left}:root .has-text-align-right{text-align:right}.has-fit-text{white-space:nowrap!important}#end-resizable-editor-section{display:none}.aligncenter{clear:both}.items-justified-left{justify-content:flex-start}.items-justified-center{justify-content:center}.items-justified-right{justify-content:flex-end}.items-justified-space-between{justify-content:space-between}.screen-reader-text{word-wrap:normal!important;border:0;clip-path:inset(50%);height:1px;margin:-1px;overflow:hidden;padding:0;position:absolute;width:1px}.screen-reader-text:focus{background-color:#ddd;clip-path:none;color:#444;display:block;font-size:1em;height:auto;left:5px;line-height:normal;padding:15px 23px 14px;text-decoration:none;top:5px;width:auto;z-index:100000}html :where(.has-border-color){border-style:solid}html :where([style*=border-color]){border-style:solid}html :where([style*=border-top-color]){border-top-style:solid}html :where([style*=border-right-color]){border-right-style:solid}html :where([style*=border-bottom-color]){border-bottom-style:solid}html :where([style*=border-left-color]){border-left-style:solid}html :where([style*=border-width]){border-style:solid}html :where([style*=border-top-width]){border-top-style:solid}html :where([style*=border-right-width]){border-right-style:solid}html :where([style*=border-bottom-width]){border-bottom-style:solid}html :where([style*=border-left-width]){border-left-style:solid}html :where(img[class*=wp-image-]){height:auto;max-width:100%}:where(figure){margin:0 0 1em}html :where(.is-position-sticky){--wp-admin--admin-bar--position-offset:var(--wp-admin--admin-bar--height,0px)}@media screen and (max-width:600px){html :where(.is-position-sticky){--wp-admin--admin-bar--position-offset:0px}}
/*# sourceURL=/wp-includes/css/dist/block-library/common.min.css */
</style>
<style id="wp-block-button-inline-css">
.wp-block-button__link{align-content:center;box-sizing:border-box;cursor:pointer;display:inline-block;height:100%;text-align:center;word-break:break-word}.wp-block-button__link.aligncenter{text-align:center}.wp-block-button__link.alignright{text-align:right}:where(.wp-block-button__link){border-radius:9999px;box-shadow:none;padding:calc(.667em + 2px) calc(1.333em + 2px);text-decoration:none}.wp-block-button[style*=text-decoration] .wp-block-button__link{text-decoration:inherit}.wp-block-buttons>.wp-block-button.has-custom-width{max-width:none}.wp-block-buttons>.wp-block-button.has-custom-width .wp-block-button__link{width:100%}.wp-block-buttons>.wp-block-button.has-custom-font-size .wp-block-button__link{font-size:inherit}.wp-block-buttons>.wp-block-button.wp-block-button__width-25{width:calc(25% - var(--wp--style--block-gap, .5em)*.75)}.wp-block-buttons>.wp-block-button.wp-block-button__width-50{width:calc(50% - var(--wp--style--block-gap, .5em)*.5)}.wp-block-buttons>.wp-block-button.wp-block-button__width-75{width:calc(75% - var(--wp--style--block-gap, .5em)*.25)}.wp-block-buttons>.wp-block-button.wp-block-button__width-100{flex-basis:100%;width:100%}.wp-block-buttons.is-vertical>.wp-block-button.wp-block-button__width-25{width:25%}.wp-block-buttons.is-vertical>.wp-block-button.wp-block-button__width-50{width:50%}.wp-block-buttons.is-vertical>.wp-block-button.wp-block-button__width-75{width:75%}.wp-block-button.is-style-squared,.wp-block-button__link.wp-block-button.is-style-squared{border-radius:0}.wp-block-button.no-border-radius,.wp-block-button__link.no-border-radius{border-radius:0!important}:root :where(.wp-block-button .wp-block-button__link.is-style-outline),:root :where(.wp-block-button.is-style-outline>.wp-block-button__link){border:2px solid;padding:.667em 1.333em}:root :where(.wp-block-button .wp-block-button__link.is-style-outline:not(.has-text-color)),:root :where(.wp-block-button.is-style-outline>.wp-block-button__link:not(.has-text-color)){color:currentColor}:root :where(.wp-block-button .wp-block-button__link.is-style-outline:not(.has-background)),:root :where(.wp-block-button.is-style-outline>.wp-block-button__link:not(.has-background)){background-color:initial;background-image:none}
/*# sourceURL=https://www.microsoft.com/en-us/research/wp-includes/blocks/button/style.min.css */
</style>
<style id="wp-block-cover-inline-css">
.wp-block-cover,.wp-block-cover-image{align-items:center;background-position:50%;box-sizing:border-box;display:flex;justify-content:center;min-height:430px;overflow:hidden;overflow:clip;padding:1em;position:relative}.wp-block-cover .has-background-dim:not([class*=-background-color]),.wp-block-cover-image .has-background-dim:not([class*=-background-color]),.wp-block-cover-image.has-background-dim:not([class*=-background-color]),.wp-block-cover.has-background-dim:not([class*=-background-color]){background-color:#000}.wp-block-cover .has-background-dim.has-background-gradient,.wp-block-cover-image .has-background-dim.has-background-gradient{background-color:initial}.wp-block-cover-image.has-background-dim:before,.wp-block-cover.has-background-dim:before{background-color:inherit;content:""}.wp-block-cover .wp-block-cover__background,.wp-block-cover .wp-block-cover__gradient-background,.wp-block-cover-image .wp-block-cover__background,.wp-block-cover-image .wp-block-cover__gradient-background,.wp-block-cover-image.has-background-dim:not(.has-background-gradient):before,.wp-block-cover.has-background-dim:not(.has-background-gradient):before{bottom:0;left:0;opacity:.5;position:absolute;right:0;top:0}.wp-block-cover-image.has-background-dim.has-background-dim-10 .wp-block-cover__background,.wp-block-cover-image.has-background-dim.has-background-dim-10 .wp-block-cover__gradient-background,.wp-block-cover-image.has-background-dim.has-background-dim-10:not(.has-background-gradient):before,.wp-block-cover.has-background-dim.has-background-dim-10 .wp-block-cover__background,.wp-block-cover.has-background-dim.has-background-dim-10 .wp-block-cover__gradient-background,.wp-block-cover.has-background-dim.has-background-dim-10:not(.has-background-gradient):before{opacity:.1}.wp-block-cover-image.has-background-dim.has-background-dim-20 .wp-block-cover__background,.wp-block-cover-image.has-background-dim.has-background-dim-20 .wp-block-cover__gradient-background,.wp-block-cover-image.has-background-dim.has-background-dim-20:not(.has-background-gradient):before,.wp-block-cover.has-background-dim.has-background-dim-20 .wp-block-cover__background,.wp-block-cover.has-background-dim.has-background-dim-20 .wp-block-cover__gradient-background,.wp-block-cover.has-background-dim.has-background-dim-20:not(.has-background-gradient):before{opacity:.2}.wp-block-cover-image.has-background-dim.has-background-dim-30 .wp-block-cover__background,.wp-block-cover-image.has-background-dim.has-background-dim-30 .wp-block-cover__gradient-background,.wp-block-cover-image.has-background-dim.has-background-dim-30:not(.has-background-gradient):before,.wp-block-cover.has-background-dim.has-background-dim-30 .wp-block-cover__background,.wp-block-cover.has-background-dim.has-background-dim-30 .wp-block-cover__gradient-background,.wp-block-cover.has-background-dim.has-background-dim-30:not(.has-background-gradient):before{opacity:.3}.wp-block-cover-image.has-background-dim.has-background-dim-40 .wp-block-cover__background,.wp-block-cover-image.has-background-dim.has-background-dim-40 .wp-block-cover__gradient-background,.wp-block-cover-image.has-background-dim.has-background-dim-40:not(.has-background-gradient):before,.wp-block-cover.has-background-dim.has-background-dim-40 .wp-block-cover__background,.wp-block-cover.has-background-dim.has-background-dim-40 .wp-block-cover__gradient-background,.wp-block-cover.has-background-dim.has-background-dim-40:not(.has-background-gradient):before{opacity:.4}.wp-block-cover-image.has-background-dim.has-background-dim-50 .wp-block-cover__background,.wp-block-cover-image.has-background-dim.has-background-dim-50 .wp-block-cover__gradient-background,.wp-block-cover-image.has-background-dim.has-background-dim-50:not(.has-background-gradient):before,.wp-block-cover.has-background-dim.has-background-dim-50 .wp-block-cover__background,.wp-block-cover.has-background-dim.has-background-dim-50 .wp-block-cover__gradient-background,.wp-block-cover.has-background-dim.has-background-dim-50:not(.has-background-gradient):before{opacity:.5}.wp-block-cover-image.has-background-dim.has-background-dim-60 .wp-block-cover__background,.wp-block-cover-image.has-background-dim.has-background-dim-60 .wp-block-cover__gradient-background,.wp-block-cover-image.has-background-dim.has-background-dim-60:not(.has-background-gradient):before,.wp-block-cover.has-background-dim.has-background-dim-60 .wp-block-cover__background,.wp-block-cover.has-background-dim.has-background-dim-60 .wp-block-cover__gradient-background,.wp-block-cover.has-background-dim.has-background-dim-60:not(.has-background-gradient):before{opacity:.6}.wp-block-cover-image.has-background-dim.has-background-dim-70 .wp-block-cover__background,.wp-block-cover-image.has-background-dim.has-background-dim-70 .wp-block-cover__gradient-background,.wp-block-cover-image.has-background-dim.has-background-dim-70:not(.has-background-gradient):before,.wp-block-cover.has-background-dim.has-background-dim-70 .wp-block-cover__background,.wp-block-cover.has-background-dim.has-background-dim-70 .wp-block-cover__gradient-background,.wp-block-cover.has-background-dim.has-background-dim-70:not(.has-background-gradient):before{opacity:.7}.wp-block-cover-image.has-background-dim.has-background-dim-80 .wp-block-cover__background,.wp-block-cover-image.has-background-dim.has-background-dim-80 .wp-block-cover__gradient-background,.wp-block-cover-image.has-background-dim.has-background-dim-80:not(.has-background-gradient):before,.wp-block-cover.has-background-dim.has-background-dim-80 .wp-block-cover__background,.wp-block-cover.has-background-dim.has-background-dim-80 .wp-block-cover__gradient-background,.wp-block-cover.has-background-dim.has-background-dim-80:not(.has-background-gradient):before{opacity:.8}.wp-block-cover-image.has-background-dim.has-background-dim-90 .wp-block-cover__background,.wp-block-cover-image.has-background-dim.has-background-dim-90 .wp-block-cover__gradient-background,.wp-block-cover-image.has-background-dim.has-background-dim-90:not(.has-background-gradient):before,.wp-block-cover.has-background-dim.has-background-dim-90 .wp-block-cover__background,.wp-block-cover.has-background-dim.has-background-dim-90 .wp-block-cover__gradient-background,.wp-block-cover.has-background-dim.has-background-dim-90:not(.has-background-gradient):before{opacity:.9}.wp-block-cover-image.has-background-dim.has-background-dim-100 .wp-block-cover__background,.wp-block-cover-image.has-background-dim.has-background-dim-100 .wp-block-cover__gradient-background,.wp-block-cover-image.has-background-dim.has-background-dim-100:not(.has-background-gradient):before,.wp-block-cover.has-background-dim.has-background-dim-100 .wp-block-cover__background,.wp-block-cover.has-background-dim.has-background-dim-100 .wp-block-cover__gradient-background,.wp-block-cover.has-background-dim.has-background-dim-100:not(.has-background-gradient):before{opacity:1}.wp-block-cover .wp-block-cover__background.has-background-dim.has-background-dim-0,.wp-block-cover .wp-block-cover__gradient-background.has-background-dim.has-background-dim-0,.wp-block-cover-image .wp-block-cover__background.has-background-dim.has-background-dim-0,.wp-block-cover-image .wp-block-cover__gradient-background.has-background-dim.has-background-dim-0{opacity:0}.wp-block-cover .wp-block-cover__background.has-background-dim.has-background-dim-10,.wp-block-cover .wp-block-cover__gradient-background.has-background-dim.has-background-dim-10,.wp-block-cover-image .wp-block-cover__background.has-background-dim.has-background-dim-10,.wp-block-cover-image .wp-block-cover__gradient-background.has-background-dim.has-background-dim-10{opacity:.1}.wp-block-cover .wp-block-cover__background.has-background-dim.has-background-dim-20,.wp-block-cover .wp-block-cover__gradient-background.has-background-dim.has-background-dim-20,.wp-block-cover-image .wp-block-cover__background.has-background-dim.has-background-dim-20,.wp-block-cover-image .wp-block-cover__gradient-background.has-background-dim.has-background-dim-20{opacity:.2}.wp-block-cover .wp-block-cover__background.has-background-dim.has-background-dim-30,.wp-block-cover .wp-block-cover__gradient-background.has-background-dim.has-background-dim-30,.wp-block-cover-image .wp-block-cover__background.has-background-dim.has-background-dim-30,.wp-block-cover-image .wp-block-cover__gradient-background.has-background-dim.has-background-dim-30{opacity:.3}.wp-block-cover .wp-block-cover__background.has-background-dim.has-background-dim-40,.wp-block-cover .wp-block-cover__gradient-background.has-background-dim.has-background-dim-40,.wp-block-cover-image .wp-block-cover__background.has-background-dim.has-background-dim-40,.wp-block-cover-image .wp-block-cover__gradient-background.has-background-dim.has-background-dim-40{opacity:.4}.wp-block-cover .wp-block-cover__background.has-background-dim.has-background-dim-50,.wp-block-cover .wp-block-cover__gradient-background.has-background-dim.has-background-dim-50,.wp-block-cover-image .wp-block-cover__background.has-background-dim.has-background-dim-50,.wp-block-cover-image .wp-block-cover__gradient-background.has-background-dim.has-background-dim-50{opacity:.5}.wp-block-cover .wp-block-cover__background.has-background-dim.has-background-dim-60,.wp-block-cover .wp-block-cover__gradient-background.has-background-dim.has-background-dim-60,.wp-block-cover-image .wp-block-cover__background.has-background-dim.has-background-dim-60,.wp-block-cover-image .wp-block-cover__gradient-background.has-background-dim.has-background-dim-60{opacity:.6}.wp-block-cover .wp-block-cover__background.has-background-dim.has-background-dim-70,.wp-block-cover .wp-block-cover__gradient-background.has-background-dim.has-background-dim-70,.wp-block-cover-image .wp-block-cover__background.has-background-dim.has-background-dim-70,.wp-block-cover-image .wp-block-cover__gradient-background.has-background-dim.has-background-dim-70{opacity:.7}.wp-block-cover .wp-block-cover__background.has-background-dim.has-background-dim-80,.wp-block-cover .wp-block-cover__gradient-background.has-background-dim.has-background-dim-80,.wp-block-cover-image .wp-block-cover__background.has-background-dim.has-background-dim-80,.wp-block-cover-image .wp-block-cover__gradient-background.has-background-dim.has-background-dim-80{opacity:.8}.wp-block-cover .wp-block-cover__background.has-background-dim.has-background-dim-90,.wp-block-cover .wp-block-cover__gradient-background.has-background-dim.has-background-dim-90,.wp-block-cover-image .wp-block-cover__background.has-background-dim.has-background-dim-90,.wp-block-cover-image .wp-block-cover__gradient-background.has-background-dim.has-background-dim-90{opacity:.9}.wp-block-cover .wp-block-cover__background.has-background-dim.has-background-dim-100,.wp-block-cover .wp-block-cover__gradient-background.has-background-dim.has-background-dim-100,.wp-block-cover-image .wp-block-cover__background.has-background-dim.has-background-dim-100,.wp-block-cover-image .wp-block-cover__gradient-background.has-background-dim.has-background-dim-100{opacity:1}.wp-block-cover-image.alignleft,.wp-block-cover-image.alignright,.wp-block-cover.alignleft,.wp-block-cover.alignright{max-width:420px;width:100%}.wp-block-cover-image.aligncenter,.wp-block-cover-image.alignleft,.wp-block-cover-image.alignright,.wp-block-cover.aligncenter,.wp-block-cover.alignleft,.wp-block-cover.alignright{display:flex}.wp-block-cover .wp-block-cover__inner-container,.wp-block-cover-image .wp-block-cover__inner-container{color:inherit;position:relative;width:100%}.wp-block-cover-image.is-position-top-left,.wp-block-cover.is-position-top-left{align-items:flex-start;justify-content:flex-start}.wp-block-cover-image.is-position-top-center,.wp-block-cover.is-position-top-center{align-items:flex-start;justify-content:center}.wp-block-cover-image.is-position-top-right,.wp-block-cover.is-position-top-right{align-items:flex-start;justify-content:flex-end}.wp-block-cover-image.is-position-center-left,.wp-block-cover.is-position-center-left{align-items:center;justify-content:flex-start}.wp-block-cover-image.is-position-center-center,.wp-block-cover.is-position-center-center{align-items:center;justify-content:center}.wp-block-cover-image.is-position-center-right,.wp-block-cover.is-position-center-right{align-items:center;justify-content:flex-end}.wp-block-cover-image.is-position-bottom-left,.wp-block-cover.is-position-bottom-left{align-items:flex-end;justify-content:flex-start}.wp-block-cover-image.is-position-bottom-center,.wp-block-cover.is-position-bottom-center{align-items:flex-end;justify-content:center}.wp-block-cover-image.is-position-bottom-right,.wp-block-cover.is-position-bottom-right{align-items:flex-end;justify-content:flex-end}.wp-block-cover-image.has-custom-content-position.has-custom-content-position .wp-block-cover__inner-container,.wp-block-cover.has-custom-content-position.has-custom-content-position .wp-block-cover__inner-container{margin:0}.wp-block-cover-image.has-custom-content-position.has-custom-content-position.is-position-bottom-left .wp-block-cover__inner-container,.wp-block-cover-image.has-custom-content-position.has-custom-content-position.is-position-bottom-right .wp-block-cover__inner-container,.wp-block-cover-image.has-custom-content-position.has-custom-content-position.is-position-center-left .wp-block-cover__inner-container,.wp-block-cover-image.has-custom-content-position.has-custom-content-position.is-position-center-right .wp-block-cover__inner-container,.wp-block-cover-image.has-custom-content-position.has-custom-content-position.is-position-top-left .wp-block-cover__inner-container,.wp-block-cover-image.has-custom-content-position.has-custom-content-position.is-position-top-right .wp-block-cover__inner-container,.wp-block-cover.has-custom-content-position.has-custom-content-position.is-position-bottom-left .wp-block-cover__inner-container,.wp-block-cover.has-custom-content-position.has-custom-content-position.is-position-bottom-right .wp-block-cover__inner-container,.wp-block-cover.has-custom-content-position.has-custom-content-position.is-position-center-left .wp-block-cover__inner-container,.wp-block-cover.has-custom-content-position.has-custom-content-position.is-position-center-right .wp-block-cover__inner-container,.wp-block-cover.has-custom-content-position.has-custom-content-position.is-position-top-left .wp-block-cover__inner-container,.wp-block-cover.has-custom-content-position.has-custom-content-position.is-position-top-right .wp-block-cover__inner-container{margin:0;width:auto}.wp-block-cover .wp-block-cover__image-background,.wp-block-cover video.wp-block-cover__video-background,.wp-block-cover-image .wp-block-cover__image-background,.wp-block-cover-image video.wp-block-cover__video-background{border:none;bottom:0;box-shadow:none;height:100%;left:0;margin:0;max-height:none;max-width:none;object-fit:cover;outline:none;padding:0;position:absolute;right:0;top:0;width:100%}.wp-block-cover .wp-block-cover__embed-background,.wp-block-cover-image .wp-block-cover__embed-background{border:none;bottom:0;box-shadow:none;height:100%;left:0;margin:0;max-height:none;max-width:none;outline:none;padding:0;pointer-events:none;position:absolute;right:0;top:0;width:100%}.wp-block-cover .wp-block-cover__embed-background .wp-block-embed__wrapper,.wp-block-cover-image .wp-block-cover__embed-background .wp-block-embed__wrapper{bottom:0;height:100%;left:0;margin:0;padding:0;position:absolute;right:0;top:0;width:100%}.wp-block-cover .wp-block-cover__embed-background .wp-block-embed__wrapper iframe,.wp-block-cover .wp-block-cover__embed-background iframe,.wp-block-cover-image .wp-block-cover__embed-background .wp-block-embed__wrapper iframe,.wp-block-cover-image .wp-block-cover__embed-background iframe{height:100vh;left:50%;min-height:100%;min-width:100%;pointer-events:none;position:absolute;top:50%;transform:translate(-50%,-50%);width:100vw}.wp-block-cover-image.has-parallax,.wp-block-cover.has-parallax,.wp-block-cover__image-background.has-parallax,video.wp-block-cover__video-background.has-parallax{background-attachment:fixed;background-repeat:no-repeat;background-size:cover}@supports (-webkit-touch-callout:inherit){.wp-block-cover-image.has-parallax,.wp-block-cover.has-parallax,.wp-block-cover__image-background.has-parallax,video.wp-block-cover__video-background.has-parallax{background-attachment:scroll}}@media (prefers-reduced-motion:reduce){.wp-block-cover-image.has-parallax,.wp-block-cover.has-parallax,.wp-block-cover__image-background.has-parallax,video.wp-block-cover__video-background.has-parallax{background-attachment:scroll}}.wp-block-cover-image.is-repeated,.wp-block-cover.is-repeated,.wp-block-cover__image-background.is-repeated,video.wp-block-cover__video-background.is-repeated{background-repeat:repeat;background-size:auto}.wp-block-cover-image-text,.wp-block-cover-image-text a,.wp-block-cover-image-text a:active,.wp-block-cover-image-text a:focus,.wp-block-cover-image-text a:hover,.wp-block-cover-text,.wp-block-cover-text a,.wp-block-cover-text a:active,.wp-block-cover-text a:focus,.wp-block-cover-text a:hover,section.wp-block-cover-image h2,section.wp-block-cover-image h2 a,section.wp-block-cover-image h2 a:active,section.wp-block-cover-image h2 a:focus,section.wp-block-cover-image h2 a:hover{color:#fff}.wp-block-cover-image .wp-block-cover.has-left-content{justify-content:flex-start}.wp-block-cover-image .wp-block-cover.has-right-content{justify-content:flex-end}.wp-block-cover-image.has-left-content .wp-block-cover-image-text,.wp-block-cover.has-left-content .wp-block-cover-text,section.wp-block-cover-image.has-left-content>h2{margin-left:0;text-align:left}.wp-block-cover-image.has-right-content .wp-block-cover-image-text,.wp-block-cover.has-right-content .wp-block-cover-text,section.wp-block-cover-image.has-right-content>h2{margin-right:0;text-align:right}.wp-block-cover .wp-block-cover-text,.wp-block-cover-image .wp-block-cover-image-text,section.wp-block-cover-image>h2{font-size:2em;line-height:1.25;margin-bottom:0;max-width:840px;padding:.44em;text-align:center;z-index:1}:where(.wp-block-cover-image:not(.has-text-color)),:where(.wp-block-cover:not(.has-text-color)){color:#fff}:where(.wp-block-cover-image.is-light:not(.has-text-color)),:where(.wp-block-cover.is-light:not(.has-text-color)){color:#000}:root :where(.wp-block-cover h1:not(.has-text-color)),:root :where(.wp-block-cover h2:not(.has-text-color)),:root :where(.wp-block-cover h3:not(.has-text-color)),:root :where(.wp-block-cover h4:not(.has-text-color)),:root :where(.wp-block-cover h5:not(.has-text-color)),:root :where(.wp-block-cover h6:not(.has-text-color)),:root :where(.wp-block-cover p:not(.has-text-color)){color:inherit}body:not(.editor-styles-wrapper) .wp-block-cover:not(.wp-block-cover:has(.wp-block-cover__background+.wp-block-cover__inner-container)) .wp-block-cover__embed-background,body:not(.editor-styles-wrapper) .wp-block-cover:not(.wp-block-cover:has(.wp-block-cover__background+.wp-block-cover__inner-container)) .wp-block-cover__image-background,body:not(.editor-styles-wrapper) .wp-block-cover:not(.wp-block-cover:has(.wp-block-cover__background+.wp-block-cover__inner-container)) .wp-block-cover__video-background{z-index:0}body:not(.editor-styles-wrapper) .wp-block-cover:not(.wp-block-cover:has(.wp-block-cover__background+.wp-block-cover__inner-container)) .wp-block-cover__background,body:not(.editor-styles-wrapper) .wp-block-cover:not(.wp-block-cover:has(.wp-block-cover__background+.wp-block-cover__inner-container)) .wp-block-cover__gradient-background,body:not(.editor-styles-wrapper) .wp-block-cover:not(.wp-block-cover:has(.wp-block-cover__background+.wp-block-cover__inner-container)) .wp-block-cover__inner-container,body:not(.editor-styles-wrapper) .wp-block-cover:not(.wp-block-cover:has(.wp-block-cover__background+.wp-block-cover__inner-container)).has-background-dim:not(.has-background-gradient):before{z-index:1}.has-modal-open body:not(.editor-styles-wrapper) .wp-block-cover:not(.wp-block-cover:has(.wp-block-cover__background+.wp-block-cover__inner-container)) .wp-block-cover__inner-container{z-index:auto}
/*# sourceURL=https://www.microsoft.com/en-us/research/wp-includes/blocks/cover/style.min.css */
</style>
<style id="wp-block-heading-inline-css">
h1:where(.wp-block-heading).has-background,h2:where(.wp-block-heading).has-background,h3:where(.wp-block-heading).has-background,h4:where(.wp-block-heading).has-background,h5:where(.wp-block-heading).has-background,h6:where(.wp-block-heading).has-background{padding:1.25em 2.375em}h1.has-text-align-left[style*=writing-mode]:where([style*=vertical-lr]),h1.has-text-align-right[style*=writing-mode]:where([style*=vertical-rl]),h2.has-text-align-left[style*=writing-mode]:where([style*=vertical-lr]),h2.has-text-align-right[style*=writing-mode]:where([style*=vertical-rl]),h3.has-text-align-left[style*=writing-mode]:where([style*=vertical-lr]),h3.has-text-align-right[style*=writing-mode]:where([style*=vertical-rl]),h4.has-text-align-left[style*=writing-mode]:where([style*=vertical-lr]),h4.has-text-align-right[style*=writing-mode]:where([style*=vertical-rl]),h5.has-text-align-left[style*=writing-mode]:where([style*=vertical-lr]),h5.has-text-align-right[style*=writing-mode]:where([style*=vertical-rl]),h6.has-text-align-left[style*=writing-mode]:where([style*=vertical-lr]),h6.has-text-align-right[style*=writing-mode]:where([style*=vertical-rl]){rotate:180deg}
/*# sourceURL=https://www.microsoft.com/en-us/research/wp-includes/blocks/heading/style.min.css */
</style>
<style id="wp-block-list-inline-css">
ol,ul{box-sizing:border-box}:root :where(.wp-block-list.has-background){padding:1.25em 2.375em}
/*# sourceURL=https://www.microsoft.com/en-us/research/wp-includes/blocks/list/style.min.css */
</style>
<style id="wp-block-paragraph-inline-css">
.is-small-text{font-size:.875em}.is-regular-text{font-size:1em}.is-large-text{font-size:2.25em}.is-larger-text{font-size:3em}.has-drop-cap:not(:focus):first-letter{float:left;font-size:8.4em;font-style:normal;font-weight:100;line-height:.68;margin:.05em .1em 0 0;text-transform:uppercase}body.rtl .has-drop-cap:not(:focus):first-letter{float:none;margin-left:.1em}p.has-drop-cap.has-background{overflow:hidden}:root :where(p.has-background){padding:1.25em 2.375em}:where(p.has-text-color:not(.has-link-color)) a{color:inherit}p.has-text-align-left[style*="writing-mode:vertical-lr"],p.has-text-align-right[style*="writing-mode:vertical-rl"]{rotate:180deg}
/*# sourceURL=https://www.microsoft.com/en-us/research/wp-includes/blocks/paragraph/style.min.css */
</style>
<style id="wp-block-buttons-inline-css">
.wp-block-buttons{box-sizing:border-box}.wp-block-buttons.is-vertical{flex-direction:column}.wp-block-buttons.is-vertical>.wp-block-button:last-child{margin-bottom:0}.wp-block-buttons>.wp-block-button{display:inline-block;margin:0}.wp-block-buttons.is-content-justification-left{justify-content:flex-start}.wp-block-buttons.is-content-justification-left.is-vertical{align-items:flex-start}.wp-block-buttons.is-content-justification-center{justify-content:center}.wp-block-buttons.is-content-justification-center.is-vertical{align-items:center}.wp-block-buttons.is-content-justification-right{justify-content:flex-end}.wp-block-buttons.is-content-justification-right.is-vertical{align-items:flex-end}.wp-block-buttons.is-content-justification-space-between{justify-content:space-between}.wp-block-buttons.aligncenter{text-align:center}.wp-block-buttons:not(.is-content-justification-space-between,.is-content-justification-right,.is-content-justification-left,.is-content-justification-center) .wp-block-button.aligncenter{margin-left:auto;margin-right:auto;width:100%}.wp-block-buttons[style*=text-decoration] .wp-block-button,.wp-block-buttons[style*=text-decoration] .wp-block-button__link{text-decoration:inherit}.wp-block-buttons.has-custom-font-size .wp-block-button__link{font-size:inherit}.wp-block-buttons .wp-block-button__link{width:100%}.wp-block-button.aligncenter{text-align:center}
/*# sourceURL=https://www.microsoft.com/en-us/research/wp-includes/blocks/buttons/style.min.css */
</style>
<style id="wp-block-spacer-inline-css">
.wp-block-spacer{clear:both}
/*# sourceURL=https://www.microsoft.com/en-us/research/wp-includes/blocks/spacer/style.min.css */
</style>
<style id="wp-block-table-inline-css">
.wp-block-table{overflow-x:auto}.wp-block-table table{border-collapse:collapse;width:100%}.wp-block-table thead{border-bottom:3px solid}.wp-block-table tfoot{border-top:3px solid}.wp-block-table td,.wp-block-table th{border:1px solid;padding:.5em}.wp-block-table .has-fixed-layout{table-layout:fixed;width:100%}.wp-block-table .has-fixed-layout td,.wp-block-table .has-fixed-layout th{word-break:break-word}.wp-block-table.aligncenter,.wp-block-table.alignleft,.wp-block-table.alignright{display:table;width:auto}.wp-block-table.aligncenter td,.wp-block-table.aligncenter th,.wp-block-table.alignleft td,.wp-block-table.alignleft th,.wp-block-table.alignright td,.wp-block-table.alignright th{word-break:break-word}.wp-block-table .has-subtle-light-gray-background-color{background-color:#f3f4f5}.wp-block-table .has-subtle-pale-green-background-color{background-color:#e9fbe5}.wp-block-table .has-subtle-pale-blue-background-color{background-color:#e7f5fe}.wp-block-table .has-subtle-pale-pink-background-color{background-color:#fcf0ef}.wp-block-table.is-style-stripes{background-color:initial;border-collapse:inherit;border-spacing:0}.wp-block-table.is-style-stripes tbody tr:nth-child(odd){background-color:#f0f0f0}.wp-block-table.is-style-stripes.has-subtle-light-gray-background-color tbody tr:nth-child(odd){background-color:#f3f4f5}.wp-block-table.is-style-stripes.has-subtle-pale-green-background-color tbody tr:nth-child(odd){background-color:#e9fbe5}.wp-block-table.is-style-stripes.has-subtle-pale-blue-background-color tbody tr:nth-child(odd){background-color:#e7f5fe}.wp-block-table.is-style-stripes.has-subtle-pale-pink-background-color tbody tr:nth-child(odd){background-color:#fcf0ef}.wp-block-table.is-style-stripes td,.wp-block-table.is-style-stripes th{border-color:#0000}.wp-block-table.is-style-stripes{border-bottom:1px solid #f0f0f0}.wp-block-table .has-border-color td,.wp-block-table .has-border-color th,.wp-block-table .has-border-color tr,.wp-block-table .has-border-color>*{border-color:inherit}.wp-block-table table[style*=border-top-color] tr:first-child,.wp-block-table table[style*=border-top-color] tr:first-child td,.wp-block-table table[style*=border-top-color] tr:first-child th,.wp-block-table table[style*=border-top-color]>*,.wp-block-table table[style*=border-top-color]>* td,.wp-block-table table[style*=border-top-color]>* th{border-top-color:inherit}.wp-block-table table[style*=border-top-color] tr:not(:first-child){border-top-color:initial}.wp-block-table table[style*=border-right-color] td:last-child,.wp-block-table table[style*=border-right-color] th,.wp-block-table table[style*=border-right-color] tr,.wp-block-table table[style*=border-right-color]>*{border-right-color:inherit}.wp-block-table table[style*=border-bottom-color] tr:last-child,.wp-block-table table[style*=border-bottom-color] tr:last-child td,.wp-block-table table[style*=border-bottom-color] tr:last-child th,.wp-block-table table[style*=border-bottom-color]>*,.wp-block-table table[style*=border-bottom-color]>* td,.wp-block-table table[style*=border-bottom-color]>* th{border-bottom-color:inherit}.wp-block-table table[style*=border-bottom-color] tr:not(:last-child){border-bottom-color:initial}.wp-block-table table[style*=border-left-color] td:first-child,.wp-block-table table[style*=border-left-color] th,.wp-block-table table[style*=border-left-color] tr,.wp-block-table table[style*=border-left-color]>*{border-left-color:inherit}.wp-block-table table[style*=border-style] td,.wp-block-table table[style*=border-style] th,.wp-block-table table[style*=border-style] tr,.wp-block-table table[style*=border-style]>*{border-style:inherit}.wp-block-table table[style*=border-width] td,.wp-block-table table[style*=border-width] th,.wp-block-table table[style*=border-width] tr,.wp-block-table table[style*=border-width]>*{border-style:inherit;border-width:inherit}
/*# sourceURL=https://www.microsoft.com/en-us/research/wp-includes/blocks/table/style.min.css */
</style>
<style id="global-styles-inline-css">
:root{--wp--preset--aspect-ratio--square: 1;--wp--preset--aspect-ratio--4-3: 4/3;--wp--preset--aspect-ratio--3-4: 3/4;--wp--preset--aspect-ratio--3-2: 3/2;--wp--preset--aspect-ratio--2-3: 2/3;--wp--preset--aspect-ratio--16-9: 16/9;--wp--preset--aspect-ratio--9-16: 9/16;--wp--preset--color--black: #171717;--wp--preset--color--cyan-bluish-gray: #abb8c3;--wp--preset--color--white: #ffffff;--wp--preset--color--pale-pink: #f78da7;--wp--preset--color--vivid-red: #cf2e2e;--wp--preset--color--luminous-vivid-orange: #ff6900;--wp--preset--color--luminous-vivid-amber: #fcb900;--wp--preset--color--light-green-cyan: #7bdcb5;--wp--preset--color--vivid-green-cyan: #00d084;--wp--preset--color--pale-cyan-blue: #8ed1fc;--wp--preset--color--vivid-cyan-blue: #0693e3;--wp--preset--color--vivid-purple: #9b51e0;--wp--preset--color--blue: #0072cc;--wp--preset--color--purple: #5c2d91;--wp--preset--color--magenta: #b4009e;--wp--preset--color--red: #e81123;--wp--preset--color--orange: #d83b01;--wp--preset--color--yellow: #ffb900;--wp--preset--color--green: #107c10;--wp--preset--color--teal: #008272;--wp--preset--color--dark-gray: #2f2f2f;--wp--preset--color--gray: #767676;--wp--preset--color--light-gray: #e3e3e3;--wp--preset--color--lighter-gray: #f2f2f2;--wp--preset--color--light-blue: #ecf8fe;--wp--preset--color--cyan-blue: #3aa0fa;--wp--preset--color--transparent: transparent;--wp--preset--gradient--vivid-cyan-blue-to-vivid-purple: linear-gradient(135deg,rgb(6,147,227) 0%,rgb(155,81,224) 100%);--wp--preset--gradient--light-green-cyan-to-vivid-green-cyan: linear-gradient(135deg,rgb(122,220,180) 0%,rgb(0,208,130) 100%);--wp--preset--gradient--luminous-vivid-amber-to-luminous-vivid-orange: linear-gradient(135deg,rgb(252,185,0) 0%,rgb(255,105,0) 100%);--wp--preset--gradient--luminous-vivid-orange-to-vivid-red: linear-gradient(135deg,rgb(255,105,0) 0%,rgb(207,46,46) 100%);--wp--preset--gradient--very-light-gray-to-cyan-bluish-gray: linear-gradient(135deg,rgb(238,238,238) 0%,rgb(169,184,195) 100%);--wp--preset--gradient--cool-to-warm-spectrum: linear-gradient(135deg,rgb(74,234,220) 0%,rgb(151,120,209) 20%,rgb(207,42,186) 40%,rgb(238,44,130) 60%,rgb(251,105,98) 80%,rgb(254,248,76) 100%);--wp--preset--gradient--blush-light-purple: linear-gradient(135deg,rgb(255,206,236) 0%,rgb(152,150,240) 100%);--wp--preset--gradient--blush-bordeaux: linear-gradient(135deg,rgb(254,205,165) 0%,rgb(254,45,45) 50%,rgb(107,0,62) 100%);--wp--preset--gradient--luminous-dusk: linear-gradient(135deg,rgb(255,203,112) 0%,rgb(199,81,192) 50%,rgb(65,88,208) 100%);--wp--preset--gradient--pale-ocean: linear-gradient(135deg,rgb(255,245,203) 0%,rgb(182,227,212) 50%,rgb(51,167,181) 100%);--wp--preset--gradient--electric-grass: linear-gradient(135deg,rgb(202,248,128) 0%,rgb(113,206,126) 100%);--wp--preset--gradient--midnight: linear-gradient(135deg,rgb(2,3,129) 0%,rgb(40,116,252) 100%);--wp--preset--font-size--small: 13px;--wp--preset--font-size--medium: 20px;--wp--preset--font-size--large: 36px;--wp--preset--font-size--x-large: 42px;--wp--preset--spacing--20: 0.44rem;--wp--preset--spacing--30: 0.67rem;--wp--preset--spacing--40: 1rem;--wp--preset--spacing--50: 1.5rem;--wp--preset--spacing--60: 2.25rem;--wp--preset--spacing--70: 3.38rem;--wp--preset--spacing--80: 5.06rem;--wp--preset--shadow--natural: 6px 6px 9px rgba(0, 0, 0, 0.2);--wp--preset--shadow--deep: 12px 12px 50px rgba(0, 0, 0, 0.4);--wp--preset--shadow--sharp: 6px 6px 0px rgba(0, 0, 0, 0.2);--wp--preset--shadow--outlined: 6px 6px 0px -3px rgb(255, 255, 255), 6px 6px rgb(0, 0, 0);--wp--preset--shadow--crisp: 6px 6px 0px rgb(0, 0, 0);}.wp-block-button .wp-block-button__link{--wp--preset--color--blue: #0072cc;--wp--preset--color--cyan-blue: #3aa0fa;--wp--preset--color--black: #171717;--wp--preset--color--white: #ffffff;}:root { --wp--style--global--content-size: 1600px;--wp--style--global--wide-size: 1600px; }:where(body) { margin: 0; }.wp-site-blocks > .alignleft { float: left; margin-right: 2em; }.wp-site-blocks > .alignright { float: right; margin-left: 2em; }.wp-site-blocks > .aligncenter { justify-content: center; margin-left: auto; margin-right: auto; }:where(.is-layout-flex){gap: 0.5em;}:where(.is-layout-grid){gap: 0.5em;}.is-layout-flow > .alignleft{float: left;margin-inline-start: 0;margin-inline-end: 2em;}.is-layout-flow > .alignright{float: right;margin-inline-start: 2em;margin-inline-end: 0;}.is-layout-flow > .aligncenter{margin-left: auto !important;margin-right: auto !important;}.is-layout-constrained > .alignleft{float: left;margin-inline-start: 0;margin-inline-end: 2em;}.is-layout-constrained > .alignright{float: right;margin-inline-start: 2em;margin-inline-end: 0;}.is-layout-constrained > .aligncenter{margin-left: auto !important;margin-right: auto !important;}.is-layout-constrained > :where(:not(.alignleft):not(.alignright):not(.alignfull)){max-width: var(--wp--style--global--content-size);margin-left: auto !important;margin-right: auto !important;}.is-layout-constrained > .alignwide{max-width: var(--wp--style--global--wide-size);}body .is-layout-flex{display: flex;}.is-layout-flex{flex-wrap: wrap;align-items: center;}.is-layout-flex > :is(*, div){margin: 0;}body .is-layout-grid{display: grid;}.is-layout-grid > :is(*, div){margin: 0;}body{padding-top: 0px;padding-right: 0px;padding-bottom: 0px;padding-left: 0px;}:root :where(.wp-element-button, .wp-block-button__link){background-color: #32373c;border-width: 0;color: #fff;font-family: inherit;font-size: inherit;font-style: inherit;font-weight: inherit;letter-spacing: inherit;line-height: inherit;padding-top: calc(0.667em + 2px);padding-right: calc(1.333em + 2px);padding-bottom: calc(0.667em + 2px);padding-left: calc(1.333em + 2px);text-decoration: none;text-transform: inherit;}.has-black-color{color: var(--wp--preset--color--black) !important;}.has-cyan-bluish-gray-color{color: var(--wp--preset--color--cyan-bluish-gray) !important;}.has-white-color{color: var(--wp--preset--color--white) !important;}.has-pale-pink-color{color: var(--wp--preset--color--pale-pink) !important;}.has-vivid-red-color{color: var(--wp--preset--color--vivid-red) !important;}.has-luminous-vivid-orange-color{color: var(--wp--preset--color--luminous-vivid-orange) !important;}.has-luminous-vivid-amber-color{color: var(--wp--preset--color--luminous-vivid-amber) !important;}.has-light-green-cyan-color{color: var(--wp--preset--color--light-green-cyan) !important;}.has-vivid-green-cyan-color{color: var(--wp--preset--color--vivid-green-cyan) !important;}.has-pale-cyan-blue-color{color: var(--wp--preset--color--pale-cyan-blue) !important;}.has-vivid-cyan-blue-color{color: var(--wp--preset--color--vivid-cyan-blue) !important;}.has-vivid-purple-color{color: var(--wp--preset--color--vivid-purple) !important;}.has-blue-color{color: var(--wp--preset--color--blue) !important;}.has-purple-color{color: var(--wp--preset--color--purple) !important;}.has-magenta-color{color: var(--wp--preset--color--magenta) !important;}.has-red-color{color: var(--wp--preset--color--red) !important;}.has-orange-color{color: var(--wp--preset--color--orange) !important;}.has-yellow-color{color: var(--wp--preset--color--yellow) !important;}.has-green-color{color: var(--wp--preset--color--green) !important;}.has-teal-color{color: var(--wp--preset--color--teal) !important;}.has-dark-gray-color{color: var(--wp--preset--color--dark-gray) !important;}.has-gray-color{color: var(--wp--preset--color--gray) !important;}.has-light-gray-color{color: var(--wp--preset--color--light-gray) !important;}.has-lighter-gray-color{color: var(--wp--preset--color--lighter-gray) !important;}.has-light-blue-color{color: var(--wp--preset--color--light-blue) !important;}.has-cyan-blue-color{color: var(--wp--preset--color--cyan-blue) !important;}.has-transparent-color{color: var(--wp--preset--color--transparent) !important;}.has-black-background-color{background-color: var(--wp--preset--color--black) !important;}.has-cyan-bluish-gray-background-color{background-color: var(--wp--preset--color--cyan-bluish-gray) !important;}.has-white-background-color{background-color: var(--wp--preset--color--white) !important;}.has-pale-pink-background-color{background-color: var(--wp--preset--color--pale-pink) !important;}.has-vivid-red-background-color{background-color: var(--wp--preset--color--vivid-red) !important;}.has-luminous-vivid-orange-background-color{background-color: var(--wp--preset--color--luminous-vivid-orange) !important;}.has-luminous-vivid-amber-background-color{background-color: var(--wp--preset--color--luminous-vivid-amber) !important;}.has-light-green-cyan-background-color{background-color: var(--wp--preset--color--light-green-cyan) !important;}.has-vivid-green-cyan-background-color{background-color: var(--wp--preset--color--vivid-green-cyan) !important;}.has-pale-cyan-blue-background-color{background-color: var(--wp--preset--color--pale-cyan-blue) !important;}.has-vivid-cyan-blue-background-color{background-color: var(--wp--preset--color--vivid-cyan-blue) !important;}.has-vivid-purple-background-color{background-color: var(--wp--preset--color--vivid-purple) !important;}.has-blue-background-color{background-color: var(--wp--preset--color--blue) !important;}.has-purple-background-color{background-color: var(--wp--preset--color--purple) !important;}.has-magenta-background-color{background-color: var(--wp--preset--color--magenta) !important;}.has-red-background-color{background-color: var(--wp--preset--color--red) !important;}.has-orange-background-color{background-color: var(--wp--preset--color--orange) !important;}.has-yellow-background-color{background-color: var(--wp--preset--color--yellow) !important;}.has-green-background-color{background-color: var(--wp--preset--color--green) !important;}.has-teal-background-color{background-color: var(--wp--preset--color--teal) !important;}.has-dark-gray-background-color{background-color: var(--wp--preset--color--dark-gray) !important;}.has-gray-background-color{background-color: var(--wp--preset--color--gray) !important;}.has-light-gray-background-color{background-color: var(--wp--preset--color--light-gray) !important;}.has-lighter-gray-background-color{background-color: var(--wp--preset--color--lighter-gray) !important;}.has-light-blue-background-color{background-color: var(--wp--preset--color--light-blue) !important;}.has-cyan-blue-background-color{background-color: var(--wp--preset--color--cyan-blue) !important;}.has-transparent-background-color{background-color: var(--wp--preset--color--transparent) !important;}.has-black-border-color{border-color: var(--wp--preset--color--black) !important;}.has-cyan-bluish-gray-border-color{border-color: var(--wp--preset--color--cyan-bluish-gray) !important;}.has-white-border-color{border-color: var(--wp--preset--color--white) !important;}.has-pale-pink-border-color{border-color: var(--wp--preset--color--pale-pink) !important;}.has-vivid-red-border-color{border-color: var(--wp--preset--color--vivid-red) !important;}.has-luminous-vivid-orange-border-color{border-color: var(--wp--preset--color--luminous-vivid-orange) !important;}.has-luminous-vivid-amber-border-color{border-color: var(--wp--preset--color--luminous-vivid-amber) !important;}.has-light-green-cyan-border-color{border-color: var(--wp--preset--color--light-green-cyan) !important;}.has-vivid-green-cyan-border-color{border-color: var(--wp--preset--color--vivid-green-cyan) !important;}.has-pale-cyan-blue-border-color{border-color: var(--wp--preset--color--pale-cyan-blue) !important;}.has-vivid-cyan-blue-border-color{border-color: var(--wp--preset--color--vivid-cyan-blue) !important;}.has-vivid-purple-border-color{border-color: var(--wp--preset--color--vivid-purple) !important;}.has-blue-border-color{border-color: var(--wp--preset--color--blue) !important;}.has-purple-border-color{border-color: var(--wp--preset--color--purple) !important;}.has-magenta-border-color{border-color: var(--wp--preset--color--magenta) !important;}.has-red-border-color{border-color: var(--wp--preset--color--red) !important;}.has-orange-border-color{border-color: var(--wp--preset--color--orange) !important;}.has-yellow-border-color{border-color: var(--wp--preset--color--yellow) !important;}.has-green-border-color{border-color: var(--wp--preset--color--green) !important;}.has-teal-border-color{border-color: var(--wp--preset--color--teal) !important;}.has-dark-gray-border-color{border-color: var(--wp--preset--color--dark-gray) !important;}.has-gray-border-color{border-color: var(--wp--preset--color--gray) !important;}.has-light-gray-border-color{border-color: var(--wp--preset--color--light-gray) !important;}.has-lighter-gray-border-color{border-color: var(--wp--preset--color--lighter-gray) !important;}.has-light-blue-border-color{border-color: var(--wp--preset--color--light-blue) !important;}.has-cyan-blue-border-color{border-color: var(--wp--preset--color--cyan-blue) !important;}.has-transparent-border-color{border-color: var(--wp--preset--color--transparent) !important;}.has-vivid-cyan-blue-to-vivid-purple-gradient-background{background: var(--wp--preset--gradient--vivid-cyan-blue-to-vivid-purple) !important;}.has-light-green-cyan-to-vivid-green-cyan-gradient-background{background: var(--wp--preset--gradient--light-green-cyan-to-vivid-green-cyan) !important;}.has-luminous-vivid-amber-to-luminous-vivid-orange-gradient-background{background: var(--wp--preset--gradient--luminous-vivid-amber-to-luminous-vivid-orange) !important;}.has-luminous-vivid-orange-to-vivid-red-gradient-background{background: var(--wp--preset--gradient--luminous-vivid-orange-to-vivid-red) !important;}.has-very-light-gray-to-cyan-bluish-gray-gradient-background{background: var(--wp--preset--gradient--very-light-gray-to-cyan-bluish-gray) !important;}.has-cool-to-warm-spectrum-gradient-background{background: var(--wp--preset--gradient--cool-to-warm-spectrum) !important;}.has-blush-light-purple-gradient-background{background: var(--wp--preset--gradient--blush-light-purple) !important;}.has-blush-bordeaux-gradient-background{background: var(--wp--preset--gradient--blush-bordeaux) !important;}.has-luminous-dusk-gradient-background{background: var(--wp--preset--gradient--luminous-dusk) !important;}.has-pale-ocean-gradient-background{background: var(--wp--preset--gradient--pale-ocean) !important;}.has-electric-grass-gradient-background{background: var(--wp--preset--gradient--electric-grass) !important;}.has-midnight-gradient-background{background: var(--wp--preset--gradient--midnight) !important;}.has-small-font-size{font-size: var(--wp--preset--font-size--small) !important;}.has-medium-font-size{font-size: var(--wp--preset--font-size--medium) !important;}.has-large-font-size{font-size: var(--wp--preset--font-size--large) !important;}.has-x-large-font-size{font-size: var(--wp--preset--font-size--x-large) !important;}.wp-block-button .wp-block-button__link.has-blue-color{color: var(--wp--preset--color--blue) !important;}.wp-block-button .wp-block-button__link.has-cyan-blue-color{color: var(--wp--preset--color--cyan-blue) !important;}.wp-block-button .wp-block-button__link.has-black-color{color: var(--wp--preset--color--black) !important;}.wp-block-button .wp-block-button__link.has-white-color{color: var(--wp--preset--color--white) !important;}.wp-block-button .wp-block-button__link.has-blue-background-color{background-color: var(--wp--preset--color--blue) !important;}.wp-block-button .wp-block-button__link.has-cyan-blue-background-color{background-color: var(--wp--preset--color--cyan-blue) !important;}.wp-block-button .wp-block-button__link.has-black-background-color{background-color: var(--wp--preset--color--black) !important;}.wp-block-button .wp-block-button__link.has-white-background-color{background-color: var(--wp--preset--color--white) !important;}.wp-block-button .wp-block-button__link.has-blue-border-color{border-color: var(--wp--preset--color--blue) !important;}.wp-block-button .wp-block-button__link.has-cyan-blue-border-color{border-color: var(--wp--preset--color--cyan-blue) !important;}.wp-block-button .wp-block-button__link.has-black-border-color{border-color: var(--wp--preset--color--black) !important;}.wp-block-button .wp-block-button__link.has-white-border-color{border-color: var(--wp--preset--color--white) !important;}
/*# sourceURL=global-styles-inline-css */
</style>
<link rel='stylesheet' id='cpsh-shortcodes-css' href='https://www.microsoft.com/en-us/research/wp-content/plugins/column-shortcodes/assets/css/shortcodes.css?ver=1.0.1' media='all' />
<link rel='stylesheet' id='moray_blocks_shared_style-css' href='https://www.microsoft.com/en-us/research/wp-content/plugins/moray-blocks/dist/css/shared-style.css?ver=0.2.0' media='all' />
<link rel='stylesheet' id='moray_blocks_frontend_style-css' href='https://www.microsoft.com/en-us/research/wp-content/plugins/moray-blocks/dist/css/style.css?ver=0.2.0' media='all' />
<link rel='stylesheet' id='msr_block_library_plugin_shared-css' href='https://www.microsoft.com/en-us/research/wp-content/plugins/msr-blocks-library/dist/css/shared.css?ver=1787066359' media='all' />
<link rel='stylesheet' id='msr_block_library_plugin_frontend-css' href='https://www.microsoft.com/en-us/research/wp-content/plugins/msr-blocks-library/dist/css/frontend.css?ver=1787066359' media='all' />
<link rel='stylesheet' id='taxonomy-image-plugin-public-css' href='https://www.microsoft.com/en-us/research/wp-content/plugins/taxonomy-images/css/style.css?ver=0.9.6' media='screen' />
<link rel='stylesheet' id='ep_general_styles-css' href='https://www.microsoft.com/en-us/research/wp-content/plugins/elasticpress/dist/css/general-styles.css?ver=66295efe92a630617c00' media='all' />
<link rel='stylesheet' id='microsoft-research-moray-css' href='https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/css/microsoft-research-moray.min.css?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332' media='all' />
<link rel='stylesheet' id='wp-components-css' href='https://www.microsoft.com/en-us/research/wp-includes/css/dist/components/style.min.css?ver=7.0.4' media='all' />
<link rel='stylesheet' id='wp-preferences-css' href='https://www.microsoft.com/en-us/research/wp-includes/css/dist/preferences/style.min.css?ver=7.0.4' media='all' />
<link rel='stylesheet' id='wp-block-editor-css' href='https://www.microsoft.com/en-us/research/wp-includes/css/dist/block-editor/style.min.css?ver=7.0.4' media='all' />
<link rel='stylesheet' id='wp-reusable-blocks-css' href='https://www.microsoft.com/en-us/research/wp-includes/css/dist/reusable-blocks/style.min.css?ver=7.0.4' media='all' />
<link rel='stylesheet' id='wp-patterns-css' href='https://www.microsoft.com/en-us/research/wp-includes/css/dist/patterns/style.min.css?ver=7.0.4' media='all' />
<link rel='stylesheet' id='wp-editor-css' href='https://www.microsoft.com/en-us/research/wp-includes/css/dist/editor/style.min.css?ver=7.0.4' media='all' />
<link rel='stylesheet' id='msr_blocks-style-css-css' href='https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/css/blocks-style.min.css?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332' media='all' />
<link rel='stylesheet' id='elasticpress-autosuggest-css' href='https://www.microsoft.com/en-us/research/wp-content/plugins/elasticpress/dist/css/autosuggest-styles.css?ver=d87f34a78edccbda21b1' media='all' />
<link rel='stylesheet' id='duet-date-picker-style-css' href='https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/vendor/duet-date-picker/duet/themes/default.css?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332' media='all' />
<script id="jquery-core-js" src="https://www.microsoft.com/en-us/research/wp-includes/js/jquery/jquery.min.js?ver=3.7.1"></script>
<script id="jquery-migrate-js" src="https://www.microsoft.com/en-us/research/wp-includes/js/jquery/jquery-migrate.min.js?ver=3.4.1"></script>
<script id="oneds-tracking-js" src="https://js.monitor.azure.com/scripts/c/ms.analytics-web-3.min.js"></script>
<link rel="https://api.w.org/" href="https://www.microsoft.com/en-us/research/wp-json/" /><link rel="alternate" title="JSON" type="application/json" href="https://www.microsoft.com/en-us/research/wp-json/wp/v2/msr-event/1140163" /><link rel="EditURI" type="application/rsd+xml" title="RSD" href="https://www.microsoft.com/en-us/research/xmlrpc.php?rsd" />
<meta name="generator" content="WordPress 7.0.4" />
<link rel='shortlink' href='https://www.microsoft.com/en-us/research/?p=1140163' />
<style>
uhf-header:not(:defined) {
display: block;
height: 54px;
}
uhf-brand:not(:defined),
uhf-contextual-nav:not(:defined),
uhf-actions:not(:defined),
uhf-global-nav:not(:defined),
uhf-search:not(:defined),
uhf-mecontrol:not(:defined),
uhf-cart:not(:defined),
uhf-dropdown:not(:defined),
uhf-popout:not(:defined) {
visibility: hidden;
}
</style>
<link rel="stylesheet" href="https://uhf.microsoft.com/statics/20260814.11.18/css/style-By05NU7M.css" /><!-- Stream WordPress user activity plugin v4.2.0 -->
<link type="text/plain" rel="author" href="https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/humans.txt" /><meta name="research-area" content="Artificial intelligence; Medical, health and genomics"><link rel="prefetch" href="https://c.s-microsoft.com" /><link rel="prefetch" href="https://www.clarity.ms" /><link rel="prefetch" href="https://connect.facebook.net" /><link rel="alternate" hreflang="x-default" href="https://www.microsoft.com/en-us/research/event/microsoft-research-forum/">
<link rel="alternate" hreflang="en-us" href="https://www.microsoft.com/en-us/research/event/microsoft-research-forum/">
<!-- Facebook Pixel Code -->
<script>
function facebookTracking() {
!function(f,b,e,v,n,t,s){if(f.fbq)return;n=f.fbq=function(){n.callMethod?
n.callMethod.apply(n,arguments):n.queue.push(arguments)};if(!f._fbq)f._fbq=n;
n.push=n;n.loaded=!0;n.version='2.0';n.queue=[];t=b.createElement(e);t.async=!0;
t.src=v;s=b.getElementsByTagName(e)[0];s.parentNode.insertBefore(t,s)}(window,
document,'script','https://connect.facebook.net/en_US/fbevents.js');
fbq('init', '435868603227390');
fbq('track', 'PageView');
}
</script>
<!-- End Facebook Pixel Code -->
<!-- LinkedIn Code -->
<script type="text/javascript">
var _linkedin_data_partner_id = "7850";
function linkedinTracking(){
var s = document.getElementsByTagName("script")[0];
var b = document.createElement("script");
b.type = "text/javascript";b.async = true;
b.src = "https://snap.licdn.com/li.lms-analytics/insight.min.js";
s.parentNode.insertBefore(b, s);
}
</script>
<!-- End LinkedIn Code -->
<!-- Clarity Code -->
<script type="text/javascript">
function clarityTracking() {
(function(c,l,a,r,i,t,y){
c[a]=c[a]||function(){(c[a].q=c[a].q||[]).push(arguments)};
t=l.createElement(r);t.async=1;t.src="https://www.clarity.ms/tag/"+i;
y=l.getElementsByTagName(r)[0];y.parentNode.insertBefore(t,y);
})(window, document, "clarity", "script", "738awl9hsy");
}
</script>
<!-- End Clarity Code -->
<style id="wp-custom-css">
.wp-block-social-link-label,
.wp-block-social-link-anchor {
margin-bottom: 0;
}
/** MSR-3986 - External Links on images **/
.wp-block-image > a.msr-external-link:after, .wp-block-media-text__media > a.msr-external-link:after {
display: none;
}
/** End **/
/** MSR-4078 **/
.wp-social-link a {
padding: 0 !important;
}
.wp-social-link {
background: transparent !important;
}
/* CTA button left-align fix — remove after next deploy */
.wp-block-buttons:not(.is-content-justification-center) > .wp-block-button:first-child.is-style-cta .wp-block-button__link:not(.has-background),
.wp-block-buttons:not(.is-content-justification-center) > .wp-block-button:first-child.is-style-link .wp-block-button__link:not(.has-background) {
padding-left: 0;
}
/** End **/
/** MSR-4077 **/
.wp-block-pullquote {
padding: 2rem 0;
}
/** End **/
/** MSRA Launch **/
li.custom-control:has(#field-associated_msr_research_lab-post-1012650-1012650) {
display: none !important;
}
/** End **/
.postid-1158539 .annotations__caption:last-child {
margin-bottom: 0
}
.postid-1158539 .tab-content .annotations__list.card.p-4 {
padding-left: 0.75rem !important;
padding-bottom: 0 !important;
padding-right: 0 !important;
}
/** Articles Fix **/
body.single-msr-blog-post.has-content-parent .content-container {max-width: none;}
.content-container iframe[title="Blubrry Podcast Player"], .content-container iframe[title="Podcast Player"] {
height: 168px !important;
}
/** UHF v 1.1 Customizations **/
uhf-dropdown-column.uhf-breakpoint--desktop {
max-width: 250px;
}
/** End UHF v 1.1 Customizations **/
/** MSR-4896 UHF banner button accessibility fixes, remove once we can confirm the fix **/
uhf-promo-banner .uhf-promo-banner__action {
color:#fff !important;
}
uhf-promo-banner .uhf-promo-banner__action:hover {
color:#fff !important;
}
/** End MSR-4896 UHF banner button accessibility fixes **/
</style>
<script type="module" src="https://uhf.microsoft.com/statics/20260814.11.18/js/entry.js"></script>
<script src="https://wcpstatic.microsoft.com/mscc/lib/v2/wcp-consent.js"></script> <style id="core-block-supports-inline-css">
.wp-elements-6f57d3bc4ec2c81f9fa9647a0817bd6d a:where(:not(.wp-element-button)){color:var(--wp--preset--color--white);}.wp-elements-e67cf533dd95c85853331ecf9703385e a:where(:not(.wp-element-button)){color:var(--wp--preset--color--black);}.wp-elements-d2f5b57c698943b521515c0f3b590be0 a:where(:not(.wp-element-button)){color:var(--wp--preset--color--black);}.wp-container-core-cover-is-layout-b3d79d46 > :where(:not(.alignleft):not(.alignright):not(.alignfull)){max-width:1388px;margin-left:auto !important;margin-right:auto !important;}.wp-container-core-cover-is-layout-b3d79d46 > .alignwide{max-width:1388px;}.wp-container-core-cover-is-layout-b3d79d46 .alignfull{max-width:none;}.wp-container-core-buttons-is-layout-fe48e5de{justify-content:center;}.wp-elements-4789963789ee9fcc3e241e4dfbbf2bfc a:where(:not(.wp-element-button)){color:var(--wp--preset--color--white);}
/*# sourceURL=core-block-supports-inline-css */
</style>
</head>
<body class="wp-singular msr-event-template-default single single-msr-event postid-1140163 wp-embed-responsive wp-theme-microsoft-research-theme microsoft-uhf ">
<div id="banner" class="site-header theme-light" data-bi-aN="header">
<uhf-header locale="en-us" partnerId="MSRESEARCH" headerId="research-header-main" theme="light">
<a slot="skip-link" class="uhf-skip-link" href="" data-m='{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "SkipToMain", "cN": "Skip to content_nonnav", "ecn": "Skip to content_nonnav", "ehn": "SkipToMain"}'>Skip to main content</a>
<uhf-promo-banner
slot="promo-banner"
banner-config="[{"browser":"anaheim","title":"Maximize your points with the Microsoft Rewards extension","paragraph":"Quick access to your daily points and offers","actionLinkText":"Add it now","actionLinkAriaLabel":"Add it now","dismissText":"No, thanks","dismissAriaLabel":"No, thanks","logoUrlDarkTheme":"https://uhf.microsoft.com/images/banners/RE4mDoE.png","logoUrlLightTheme":"https://uhf.microsoft.com/images/banners/RE4mDoE.png","backgroundColorDarkTheme":"b-black","backgroundColorLightTheme":"b-white","actionLinkBackgroundColorDarkTheme":"btn-white","actionLinkBackgroundColorLightTheme":"btn-light-blue","cookieExpiration":7,"extensionType":"windows10only","extensionUrl":"https://browserdefaults.microsoft.com/extn/redirect/?xid=106\u0026channel=uhf\u0026pc=U785"},{"browser":"edge","title":"Try the browser recommended by Microsoft","paragraph":"Get speed, security and privacy with Microsoft Edge","actionLinkText":"Download now","actionLinkAriaLabel":"Download now","dismissText":"No thanks","dismissAriaLabel":"No thanks","logoUrlDarkTheme":"https://uhf.microsoft.com/images/banners/RE4xdax.png","logoUrlLightTheme":"https://uhf.microsoft.com/images/banners/RE4xdax.png","backgroundColorDarkTheme":"b-black","backgroundColorLightTheme":"b-white","actionLinkBackgroundColorDarkTheme":"btn-white","actionLinkBackgroundColorLightTheme":"btn-light-blue","cookieExpiration":30,"extensionType":"windows10only","extensionUrl":"https://aka.ms/MicrosoftEdgeDownload"},{"browser":"chrome","title":"Maximize your points with the Microsoft Rewards extension","paragraph":"Quick access to your daily points and offers","actionLinkText":"Add it now","actionLinkAriaLabel":"Add it now","dismissText":"No thanks","dismissAriaLabel":"No thanks","logoUrlDarkTheme":"https://uhf.microsoft.com/images/banners/RE4mDoE.png","logoUrlLightTheme":"https://uhf.microsoft.com/images/banners/RE4mDoE.png","backgroundColorDarkTheme":"b-black","backgroundColorLightTheme":"b-white","actionLinkBackgroundColorDarkTheme":"btn-white","actionLinkBackgroundColorLightTheme":"btn-light-blue","cookieExpiration":14,"extensionType":"windows10only","extensionUrl":"https://browserdefaults.microsoft.com/extn/redirect/?xid=106\u0026channel=uhf\u0026pc=U785"},{"browser":"firefox","title":"Maximize your points with the Microsoft Rewards extension","paragraph":"Quick access to your daily points and offers","actionLinkText":"Add it now","actionLinkAriaLabel":"Add it now","dismissText":"No thanks","dismissAriaLabel":"No thanks","logoUrlDarkTheme":"https://uhf.microsoft.com/images/banners/RE4mFZT.png","logoUrlLightTheme":"https://uhf.microsoft.com/images/banners/RE4mDoE.png","backgroundColorDarkTheme":"b-blue","backgroundColorLightTheme":"b-white","actionLinkBackgroundColorDarkTheme":"btn-white","actionLinkBackgroundColorLightTheme":"btn-blue","cookieExpiration":30,"extensionType":"rewards","extensionUrl":"https://browserdefaults.microsoft.com/extn/redirect/?xid=106\u0026channel=uhf\u0026pc=U785"}]"
></uhf-promo-banner>
<uhf-brand slot="brand">
<a href="https://www.microsoft.com" class="uhf-microsoft-logo" slot="microsoft-logo" aria-label="Microsoft" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Anchor", "cN": "GlobalNav_Logo_cont", "ecn": "GlobalNav_Logo_cont", "ehn": "Anchor"}">
<img src="https://uhf.microsoft.com/images/microsoft/RE1Mu3b.png" alt="Microsoft" />
</a>
<a href="/en-us/research/" class="uhf-site-logo" slot="brand-logo" aria-label="Research" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Anchor", "cN": "CatNav_Research_nav", "ecn": "CatNav_Research_nav", "ehn": "Anchor"}">
<span>Research</span> </a>
</uhf-brand>
<uhf-contextual-nav
slot="contextual-nav"
overflowText="More"
brand="Research"
homeUrl="/en-us/research/"
homeText="Home"
data-nav-label="Contextual menu"
theme=cat-theme-gray
>
<uhf-dropdown
text="Our research"
id=""
data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_nav", "ecn": "CatNav_OurResearch_nav", "ehn": "OurResearch"}"
>
<uhf-dropdown-column title="Resources" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_nav", "ecn": "CatNav_OurResearch_nav", "ehn": "OurResearch"}">
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/publications/" id="Publications" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Resources_Publications_nav", "ecn": "CatNav_OurResearch_Resources_Publications_nav", "ehn": "OurResearch"}">Publications</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/tools/" id="CodeDatasets" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Resources_Code & data_nav", "ecn": "CatNav_OurResearch_Resources_Code & data_nav", "ehn": "OurResearch"}">Code & data</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/people/" id="People" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Resources_People_nav", "ecn": "CatNav_OurResearch_Resources_People-resources_nav", "ehn": "OurResearch"}">People</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/blog/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Resources_Microsoft Research blog_nav", "ecn": "CatNav_OurResearch_Resources_MicrosoftResearchBlog-resources_nav", "ehn": "OurResearch"}">Microsoft Research blog</a>
</uhf-dropdown-column>
<uhf-dropdown-column title="Research areas: Intelligence" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_nav", "ecn": "CatNav_OurResearch_nav", "ehn": "OurResearch"}">
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/focus-area/ai-and-microsoft-research/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Intelligence_Artificial intelligence_nav", "ecn": "CatNav_OurResearch_Intelligence_ArtificialIntelligence_nav", "ehn": "OurResearch"}">Artificial intelligence</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/audio-acoustics/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Intelligence_Audio & acoustics_nav", "ecn": "CatNav_OurResearch_Intelligence_audioacoustics_nav", "ehn": "OurResearch"}">Audio & acoustics</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/computer-vision/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Intelligence_Computer vision_nav", "ecn": "CatNav_OurResearch_Intelligence_Computervision_nav", "ehn": "OurResearch"}">Computer vision</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/graphics-and-multimedia/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Intelligence_Graphics & multimedia_nav", "ecn": "CatNav_OurResearch_Intelligence_Graphicsmultimedia_nav", "ehn": "OurResearch"}">Graphics & multimedia</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/human-computer-interaction/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Intelligence_Human-computer interaction_nav", "ecn": "CatNav_OurResearch_Intelligence_Humancomputerinteraction_nav", "ehn": "OurResearch"}">Human-computer interaction</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/human-language-technologies/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Intelligence_Human language technologies_nav", "ecn": "CatNav_OurResearch_Intelligence_Humanlanguagetechnologies_nav", "ehn": "OurResearch"}">Human language technologies</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/search-information-retrieval/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Intelligence_Search & information retrieval_nav", "ecn": "CatNav_OurResearch_Intelligence_Searchinformationretrieval_nav", "ehn": "OurResearch"}">Search & information retrieval</a>
</uhf-dropdown-column>
<uhf-dropdown-column title="Research areas: Systems" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_nav", "ecn": "CatNav_OurResearch_nav", "ehn": "OurResearch"}">
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/data-platform-analytics/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Systems_Data platforms and analytics_nav", "ecn": "CatNav_OurResearch_Systems_Datamanagementanalysisvisualization_nav", "ehn": "OurResearch"}">Data platforms and analytics</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/hardware-devices/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Systems_Hardware & devices_nav", "ecn": "CatNav_OurResearch_Systems_Hardwaredevices_nav", "ehn": "OurResearch"}">Hardware & devices</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/programming-languages-software-engineering/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Systems_Programming languages & software engineering_nav", "ecn": "CatNav_OurResearch_Systems_Programminglanguagessoftwareengineering_nav", "ehn": "OurResearch"}">Programming languages & software engineering</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/quantum/" id="Quantum computing" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Systems_Quantum computing_nav", "ecn": "CatNav_OurResearch_Systems_QuantumComputing_nav", "ehn": "OurResearch"}">Quantum computing</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/security-privacy-cryptography/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Systems_Security, privacy & cryptography_nav", "ecn": "CatNav_OurResearch_Systems_Securityprivacycryptography_nav", "ehn": "OurResearch"}">Security, privacy & cryptography</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/systems-and-networking/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Systems_Systems & networking_nav", "ecn": "CatNav_OurResearch_Systems_Computersystemsnetworking_nav", "ehn": "OurResearch"}">Systems & networking</a>
</uhf-dropdown-column>
<uhf-dropdown-column title="Research areas: Theory" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_nav", "ecn": "CatNav_OurResearch_nav", "ehn": "OurResearch"}">
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/algorithms/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Theory_Algorithms_nav", "ecn": "CatNav_OurResearch_Theory_Algorithms_nav", "ehn": "OurResearch"}">Algorithms</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/computational-sciences-mathematics/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Theory_Mathematics_nav", "ecn": "CatNav_OurResearch_Theory_Mathematics_nav", "ehn": "OurResearch"}">Mathematics</a>
</uhf-dropdown-column>
<uhf-dropdown-column title="Research areas: Other Sciences" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_nav", "ecn": "CatNav_OurResearch_nav", "ehn": "OurResearch"}">
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/ecology-environment/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Other Sciences_Ecology & environment_nav", "ecn": "CatNav_OurResearch_Other Sciences_Ecologyenvironment_nav", "ehn": "OurResearch"}">Ecology & environment</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/economics/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Other Sciences_Economics_nav", "ecn": "CatNav_OurResearch_Other Sciences_Economics_nav", "ehn": "OurResearch"}">Economics</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/medical-health-genomics/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Other Sciences_Medical, health & genomics_nav", "ecn": "CatNav_OurResearch_Other Sciences_Medicalhealthgenomics_nav", "ehn": "OurResearch"}">Medical, health & genomics</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/social-sciences/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Other Sciences_Social sciences_nav", "ecn": "CatNav_OurResearch_Other Sciences_Socialsciences_nav", "ehn": "OurResearch"}">Social sciences</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/research-area/technology-for-emerging-markets/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Our research", "cN": "CatNav_Our research_Research areas: Other Sciences_Technology for emerging markets_nav", "ecn": "CatNav_OurResearch_Other Sciences_Technologyemergingmarkets_nav", "ehn": "OurResearch"}">Technology for emerging markets</a>
</uhf-dropdown-column>
</uhf-dropdown>
<uhf-dropdown
text="Programs & events"
id=""
data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Programs & events", "cN": "CatNav_Programs & events_nav", "ecn": "CatNav_ProgramsEvents_nav", "ehn": "ProgramsEvents"}"
>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/academic-programs/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Programs & events", "cN": "CatNav_Programs & events_Academic programs_nav", "ecn": "CatNav_ProgramsEvents_Academic programs_nav", "ehn": "ProgramsEvents"}">Academic programs</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/events-conferences/" id="Events & academic conferences" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Programs & events", "cN": "CatNav_Programs & events_Events & academic conferences_nav", "ecn": "CatNav_ProgramsEvents_Events & academic conferences_nav", "ehn": "ProgramsEvents"}">Events & academic conferences</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://researchforum.microsoft.com" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Programs & events", "cN": "CatNav_Programs & events_Microsoft Research Forum_nav", "ecn": "CatNav_ProgramsEvents_Microsoft Research Forum_nav", "ehn": "ProgramsEvents"}">Microsoft Research Forum</a>
</uhf-dropdown>
<uhf-dropdown
text="Connect & learn"
id="Connect learn"
data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Connect & learn", "cN": "CatNav_Connect & learn_nav", "ecn": "CatNav_Connect & learn_nav", "ehn": "Connect & learn"}"
>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/behind-the-tech " data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Connect & learn", "cN": "CatNav_Connect & learn_Behind the Tech podcast_nav", "ecn": "CatNav_Connect & learn_BehindtheTech_nav", "ehn": "Connect & learn"}">Behind the Tech podcast</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/blog" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Connect & learn", "cN": "CatNav_Connect & learn_Microsoft Research blog_nav", "ecn": "CatNav_Connect & learn_MicrosoftResearchblog-blog_nav", "ehn": "Connect & learn"}">Microsoft Research blog</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://researchforum.microsoft.com" id="Microsoft Research Forum" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Connect & learn", "cN": "CatNav_Connect & learn_Microsoft Research Forum_nav", "ecn": "CatNav_Connect & learn_Microsoft Research Forum_nav", "ehn": "Connect & learn"}">Microsoft Research Forum</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/podcast/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "Connect & learn", "cN": "CatNav_Connect & learn_Microsoft Research podcast_nav", "ecn": "CatNav_Connect & learn_MicrosoftResearchpodcast_nav", "ehn": "Connect & learn"}">Microsoft Research podcast</a>
</uhf-dropdown>
<uhf-dropdown
text="About"
id="About"
data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_nav", "ecn": "CatNav_About_nav", "ehn": "About"}"
>
<uhf-dropdown-column title="People & news" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_nav", "ecn": "CatNav_About_nav", "ehn": "About"}">
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/about-microsoft-research/" id="About Microsoft Research" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_People & news_About Microsoft Research_nav", "ecn": "CatNav_About_People_/en-us/research/about-microsoft-research/_nav", "ehn": "About"}">About Microsoft Research</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/careers/" id="CareersInternships" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_People & news_Careers & internships_nav", "ecn": "CatNav_About_People_CareersInternships_nav", "ehn": "About"}">Careers & internships</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/people/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_People & news_People_nav", "ecn": "CatNav_About_People_People-about_nav", "ehn": "About"}">People</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/microsoft-research-emeritus-program/" id="EmeritusProgram" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_People & news_Emeritus program_nav", "ecn": "CatNav_About_People_Emeritus program_nav", "ehn": "About"}">Emeritus program</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/news-and-awards/" id="NewsAwards" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_People & news_News & awards_nav", "ecn": "CatNav_About_People_News & awards_nav", "ehn": "About"}">News & awards</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://info.microsoft.com/ww-landing-microsoft-research-newsletter.html?wt.mc_id=S-webpage_msr-homepage" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_People & news_Microsoft Research newsletter_nav", "ecn": "CatNav_About_People_Newsletter-about_nav", "ehn": "About"}">Microsoft Research newsletter</a>
</uhf-dropdown-column>
<uhf-dropdown-column title="Microsoft Research Labs" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_nav", "ecn": "CatNav_About_nav", "ehn": "About"}">
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/microsoft-research-lab-africa-nairobi/" id="Africa" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Microsoft Research Labs_Africa_nav", "ecn": "CatNav_About_MicrosoftResearchLabs_Africa_nav", "ehn": "About"}">Africa</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/microsoft-research-ai-for-science/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Microsoft Research Labs_AI for Science_nav", "ecn": "CatNav_About_MicrosoftResearchLabs_AI for Science_nav", "ehn": "About"}">AI for Science</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/ai-frontiers/" id="AI Frontiers" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Microsoft Research Labs_AI Frontiers_nav", "ecn": "CatNav_About_MicrosoftResearchLabs_AI Frontiers_nav", "ehn": "About"}">AI Frontiers</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/microsoft-research-asia/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Microsoft Research Labs_Asia-Pacific_nav", "ecn": "CatNav_About_MicrosoftResearchLabs_AsiaPacific_nav", "ehn": "About"}">Asia-Pacific</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/microsoft-research-cambridge/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Microsoft Research Labs_Cambridge_nav", "ecn": "CatNav_About_MicrosoftResearchLabs_CambridgeLab_nav", "ehn": "About"}">Cambridge</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/microsoft-health-futures/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Microsoft Research Labs_Health Futures_nav", "ecn": "CatNav_About_MicrosoftResearchLabs_Health Futures_nav", "ehn": "About"}">Health Futures</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/microsoft-research-india/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Microsoft Research Labs_India_nav", "ecn": "CatNav_About_MicrosoftResearchLabs_IndiaLab_nav", "ehn": "About"}">India</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/microsoft-research-montreal/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Microsoft Research Labs_Montreal_nav", "ecn": "CatNav_About_MicrosoftResearchLabs_MontrealLab_nav", "ehn": "About"}">Montreal</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/microsoft-research-new-england/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Microsoft Research Labs_New England_nav", "ecn": "CatNav_About_MicrosoftResearchLabs_NewEnglandLab_nav", "ehn": "About"}">New England</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/microsoft-research-new-york/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Microsoft Research Labs_New York City_nav", "ecn": "CatNav_About_MicrosoftResearchLabs_NewYorkCityLab_nav", "ehn": "About"}">New York City</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/microsoft-research-redmond/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Microsoft Research Labs_Redmond_nav", "ecn": "CatNav_About_MicrosoftResearchLabs_RedmondLab_nav", "ehn": "About"}">Redmond</a>
</uhf-dropdown-column>
<uhf-dropdown-column title="Other labs" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_nav", "ecn": "CatNav_About_nav", "ehn": "About"}">
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/applied-sciences-group/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Other labs_Applied Sciences_nav", "ecn": "CatNav_About_Other labs_AppliedSciencesLab_nav", "ehn": "About"}">Applied Sciences</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/mixed-reality-ai-lab-cambridge/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Other labs_Mixed Reality & AI - Cambridge_nav", "ecn": "CatNav_About_Other labs_Mixed Reality & AI - Cambridge_nav", "ehn": "About"}">Mixed Reality & AI - Cambridge</a>
<a class="uhf-nav-item uhf-dropdown-link" href="/en-us/research/lab/mixed-reality-ai-zurich/" id="Mixed-reality" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "About", "cN": "CatNav_About_Other labs_Mixed Reality & AI - Zurich_nav", "ecn": "CatNav_About_Other labs_Mixed Reality & AI - Zurich_nav", "ehn": "About"}">Mixed Reality & AI - Zurich</a>
</uhf-dropdown-column>
</uhf-dropdown> <a
class="uhf-nav-item uhf-nav-cta"
href="https://researchforum.microsoft.com"
slot="CTA"
id="NewsletterButton"
data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "CatNav_OneMicrosoftBar_Newsletter-Button_nav", "ecn": "CatNav_OneMicrosoftBar_Newsletter-Button_nav", "ehn": "OneMicrosoftBar"}"
>
Register: Research Forum
</a>
</uhf-contextual-nav> <uhf-actions slot="actions">
<uhf-global-nav slot="global-nav" text="All Microsoft" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_nonnav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_nonnav", "ehn": "OneMicrosoftBar"}" data-nav-label="All Microsoft menu">
<uhf-dropdown-header slot="header">
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/security" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Microsoft Security_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Microsoft Security_nav", "ehn": "OneMicrosoftBar"}">Microsoft Security</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://azure.microsoft.com/en-us/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Azure_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Azure_nav", "ehn": "OneMicrosoftBar"}">Azure</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://dynamics.microsoft.com/en-us/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Dynamics 365_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Dynamics 365_nav", "ehn": "OneMicrosoftBar"}">Dynamics 365</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/microsoft-365/business/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Microsoft 365_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Microsoft 365_nav", "ehn": "OneMicrosoftBar"}">Microsoft 365</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/microsoft-teams/group-chat-software" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Microsoft Teams_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Microsoft Teams_nav", "ehn": "OneMicrosoftBar"}">Microsoft Teams</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/windows-365" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Windows 365_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_Windows 365_nav", "ehn": "OneMicrosoftBar"}">Windows 365</a>
</uhf-dropdown-header>
<uhf-dropdown-column title="Tech & innovation" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_nav", "ehn": "OneMicrosoftBar"}">
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/ai?icid=DSM_AllCommercial_AI" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_Microsoft AI_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_More_TechInnovation__AI_nav", "ehn": "OneMicrosoftBar"}">Microsoft AI</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://azure.microsoft.com/en-us/solutions/space/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_Azure Space_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_More_TechInnovation_AzureSpace_nav", "ehn": "OneMicrosoftBar"}">Azure Space</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/mixed-reality/windows-mixed-reality" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_Mixed reality_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_More_TechInnovation_MixedReality_nav", "ehn": "OneMicrosoftBar"}">Mixed reality</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/hololens" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_Microsoft HoloLens_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_More_TechInnovation_MicrosoftHololens_nav", "ehn": "OneMicrosoftBar"}">Microsoft HoloLens</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/microsoft-viva" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_Microsoft Viva_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_More_TechInnovation_Microsoft Viva_nav", "ehn": "OneMicrosoftBar"}">Microsoft Viva</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://azure.microsoft.com/en-us/solutions/quantum-computing/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_Quantum computing_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_More_TechInnovation_QuantumComputing_nav", "ehn": "OneMicrosoftBar"}">Quantum computing</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/corporate-responsibility/sustainability?icid=DSM_AllCommercial_Sustainability" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_Sustainability_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Tech & innovation_More_TechInnovation_Sustainability_nav", "ehn": "OneMicrosoftBar"}">Sustainability</a>
</uhf-dropdown-column>
<uhf-dropdown-column title="Industries" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_nav", "ehn": "OneMicrosoftBar"}">
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/education" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_Education_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_More_Industries_Education_nav", "ehn": "OneMicrosoftBar"}">Education</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/industry/automotive" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_Automotive_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_More_Industries_Automotive_nav", "ehn": "OneMicrosoftBar"}">Automotive</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/industry/financial-services/banking" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_Financial services_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_More_Industries_Financialservices_nav", "ehn": "OneMicrosoftBar"}">Financial services</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/industry/government" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_Government_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_More_Industries_Government_nav", "ehn": "OneMicrosoftBar"}">Government</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/industry/health/microsoft-cloud-for-healthcare" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_Healthcare_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_More_Industries_Health_nav", "ehn": "OneMicrosoftBar"}">Healthcare</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/industry/manufacturing/microsoft-cloud-for-manufacturing" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_Manufacturing_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_More_Industries_Manufacturing_nav", "ehn": "OneMicrosoftBar"}">Manufacturing</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/industry/consumer-goods" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_Retail_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Industries_More_Industries_Retail_nav", "ehn": "OneMicrosoftBar"}">Retail</a>
</uhf-dropdown-column>
<uhf-dropdown-column title="Partners" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_nav", "ehn": "OneMicrosoftBar"}">
<a class="uhf-nav-item uhf-dropdown-link" href="https://partner.microsoft.com/en-US/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Partners_Find a partner_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Partners_More_Partner_FindPartner_nav", "ehn": "OneMicrosoftBar"}">Find a partner</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://partner.microsoft.com/en-US/membership/cloud-solution-provider" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Partners_Become a partner_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Partners_More_Partner_BecomePartner_nav", "ehn": "OneMicrosoftBar"}">Become a partner</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://partner.microsoft.com/en-us/membership" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Partners_Partner Network_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Partners_More_Partner_PartnerNetwork_nav", "ehn": "OneMicrosoftBar"}">Partner Network</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://marketplace.microsoft.com?icid=DSM_AllCommercial_Marketplace&ocid=cmm3c8ee9bs" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Partners_Microsoft Marketplace_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Partners_More_Partner_Marketplace_nav", "ehn": "OneMicrosoftBar"}">Microsoft Marketplace</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/software-development-companies?icid=DSM_AllCommercial_SoftwareCompanies&ocid=cmm3c8ee9bs" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Partners_Software companies_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Partners_Software companies_nav", "ehn": "OneMicrosoftBar"}">Software companies</a>
</uhf-dropdown-column>
<uhf-dropdown-column title="Resources" show-tooltip="false" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_nav", "ehn": "OneMicrosoftBar"}">
<a class="uhf-nav-item uhf-dropdown-link" href="https://blogs.microsoft.com/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_Blog_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_More_Resources_Blog_nav", "ehn": "OneMicrosoftBar"}">Blog</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://about.ads.microsoft.com/en-us?s_cid=dig-src_uhfcomm" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_Microsoft Advertising_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_More_Resources_MicrosoftAdvertising_nav", "ehn": "OneMicrosoftBar"}">Microsoft Advertising</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://developer.microsoft.com/en-us/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_Developer Center_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_More_Resources_DeveloperCenter_nav", "ehn": "OneMicrosoftBar"}">Developer Center</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://learn.microsoft.com/docs/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_Documentation_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_More_Resources_Documentation_nav", "ehn": "OneMicrosoftBar"}">Documentation</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/events" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_Events_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_More_Resources_Events_nav", "ehn": "OneMicrosoftBar"}">Events</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/licensing/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_Licensing_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_More_Resources_Licensing_nav", "ehn": "OneMicrosoftBar"}">Licensing</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://learn.microsoft.com/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_Microsoft Learn_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_More_Resources_MicrosoftLearn_nav", "ehn": "OneMicrosoftBar"}">Microsoft Learn</a>
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/research/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalHeader", "hn": "OneMicrosoftBar", "cN": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_Microsoft Research_nav", "ecn": "GlobalNav_OneMicrosoftBar_AllMicrosoft_More_Resources_More_Resources_MicrosoftResearch_nav", "ehn": "OneMicrosoftBar"}">Microsoft Research</a>
</uhf-dropdown-column>
<uhf-dropdown-footer slot="footer">
<a class="uhf-nav-item uhf-dropdown-link" href="https://www.microsoft.com/en-us/sitemap" data-m="">View Sitemap</a>
</uhf-dropdown-footer>
</uhf-global-nav>
<uhf-search
slot="search"
placeholder="Search Microsoft Research"
search-label="Search"
cancel-label="Cancel"
suggestions-available-text="{0} suggestions available"
searchUrl="/en-us/research/search/"
autoSuggestUrl=""
queryParameterName="q"
>
</uhf-search>
</uhf-actions>
</uhf-header>
</div>
<main class="event" data-bi-aN="body" awa-sitesection="event single" id="main" role="main">
<div class="wp-block-cover is-light has-white-color has-text-color has-link-color wp-elements-6f57d3bc4ec2c81f9fa9647a0817bd6d" style="min-height:25vw;aspect-ratio:unset;"><img fetchpriority="high" decoding="async" width="1024" height="384" class="wp-block-cover__image-background wp-image-1147621 size-large" alt="Research Forum | Episode 5 - abstract background with colorful hexagons" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/08/Research-Forum_header_1920x720-final-1024x384.jpg" data-object-fit="cover" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/08/Research-Forum_header_1920x720-final-1024x384.jpg 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/08/Research-Forum_header_1920x720-final-300x113.jpg 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/08/Research-Forum_header_1920x720-final-768x288.jpg 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/08/Research-Forum_header_1920x720-final-1536x576.jpg 1536w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/08/Research-Forum_header_1920x720-final-1600x600.jpg 1600w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/08/Research-Forum_header_1920x720-final-240x90.jpg 240w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/08/Research-Forum_header_1920x720-final.jpg 1920w" sizes="(max-width: 1024px) 100vw, 1024px" /><span aria-hidden="true" class="wp-block-cover__background has-purple-background-color has-background-dim-0 has-background-dim"></span><div class="wp-block-cover__inner-container is-layout-constrained wp-container-core-cover-is-layout-b3d79d46 wp-block-cover-is-layout-constrained">
<p class="wp-block-paragraph"></p>
<h1 id="microsoft-research-forum" class="wp-block-heading has-black-color has-text-color has-link-color wp-elements-e67cf533dd95c85853331ecf9703385e">Microsoft Research Forum</h1>
<p class="has-black-color has-text-color has-link-color wp-elements-d2f5b57c698943b521515c0f3b590be0 wp-block-paragraph">Location: Virtual</p>
<div class="wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex">
<div class="wp-block-button"><a data-bi-type="button" class="wp-block-button__link wp-element-button" href="http://aka.ms/researchforum-register" target="_blank" rel="noreferrer noopener">Register for the series</a></div>
</div>
</div></div>
<div class="wp-block-msr-content-tabs container" data-bi-aN="content">
<section class="row row-cols-1" data-bi-tN="content-tabs">
<div class="col">
<nav class="nav-container" aria-label="Subpage navigation">
<ul class="nav my-4">
<li class="nav-item active">
<a
id="tab-overview-tab"
aria-label="Active page: Overview"
class="nav-link"
href="https://www.microsoft.com/en-us/research/event/microsoft-research-forum/"
data-bi-cN="Overview"
data-bi-type="tab"
data-bi-tN="content-tab"
>
Overview </a>
</li>
<li class="nav-item ">
<a
id="tab-past-episodes-tab"
class="nav-link"
href="https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/"
data-bi-cN="Past episodes"
data-bi-type="tab"
data-bi-tN="content-tab"
>
Past episodes </a>
</li>
</ul>
</nav>
</div>
<div class="col">
<div
class="tab-pane fade pl-0 col-12 active show"
id="overview"
data-bi-aN="Overview"
>
<p class="wp-block-paragraph">Microsoft Research Forum is a virtual series showcasing research with real-world impact, from fundamental exploration to responsible AI, scaling innovation through products and open source, and supporting positive change in society.</p>
<div style="height:30px" aria-hidden="true" class="wp-block-spacer"></div>
<div style="padding-bottom:64px;padding-top:64px" class="wp-block-msr-immersive-section alignfull row">
<div class="container">
<div class="wp-block-msr-immersive-section__wrapper">
<h2 class="wp-block-heading has-text-align-center" id="catch-up-on-season-2-episode-4">Catch up on Season 2, Episode 4</h2>
<p class="has-text-align-center wp-block-paragraph">Aired May 14, 2026</p>
<div style="height:20px" aria-hidden="true" class="wp-block-spacer"></div>
<div class="wp-block-msr-cards msr-cards msr-cards--default mt-4 has-text-align-left has-cards" data-bi-aN="Cards">
<div class="msr-cards__inner">
<div class="mt-4">
<ul class="my-0 list-unstyled row row-cols-1 row-cols-sm-2 row-cols-lg-3">
<li class="my-0 msr-cards__card msr-cards__card--default col">
<div class="card has-spectrum-border-top__hover material-card h-100 p-0">
<div class="embed-responsive embed-responsive-16by9">
<img decoding="async" width="380" height="214" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1-640x360.jpg" class="card-img embed-responsive-item img-object-cover" alt="Microsoft Research Forum S2E4 | Harkirat Behl, Weili Shi, Hussein Mozannar | MagenticLite" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1-640x360.jpg 640w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1-300x169.jpg 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1-1024x576.jpg 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1-768x432.jpg 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1-1066x600.jpg 1066w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1-655x368.jpg 655w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1-240x135.jpg 240w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1-960x540.jpg 960w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1.jpg 1280w" sizes="(max-width: 380px) 100vw, 380px" /> </div>
<div class="card-body p-4 pt-3">
<h3 class="h5">
<a href="https://www.microsoft.com/en-us/research/video/magenticlite-a-full-stack-agentic-experience-powered-by-small-models/" data-bi-cN="MagenticLite: A full-stack agentic experience powered by Small Models" class="text-decoration-none text-body"><span>MagenticLite: A full-stack agentic experience powered by Small Models</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span></a> </h3>
<div class="card__description card__citation small">
</div>
</div>
</div>
</li>
<li class="my-0 msr-cards__card msr-cards__card--default col">
<div class="card has-spectrum-border-top__hover material-card h-100 p-0">
<div class="embed-responsive embed-responsive-16by9">
<img decoding="async" width="380" height="214" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1-640x360.jpg" class="card-img embed-responsive-item img-object-cover" alt="Microsoft Research Forum S2E4 | Peli de Halleux | Introducing GitHub Agentic Workflows: AI that runs your repo" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1-640x360.jpg 640w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1-300x169.jpg 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1-1024x576.jpg 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1-768x432.jpg 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1-1066x600.jpg 1066w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1-655x368.jpg 655w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1-240x135.jpg 240w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1-960x540.jpg 960w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1.jpg 1280w" sizes="(max-width: 380px) 100vw, 380px" /> </div>
<div class="card-body p-4 pt-3">
<h3 class="h5">
<a href="https://www.microsoft.com/en-us/research/video/introducing-github-agentic-workflows-ai-that-runs-your-repo/" data-bi-cN="Introducing GitHub Agentic Workflows: AI that runs your repo" class="text-decoration-none text-body"><span>Introducing GitHub Agentic Workflows: AI that runs your repo</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span></a> </h3>
<div class="card__description card__citation small">
</div>
</div>
</div>
</li>
<li class="my-0 msr-cards__card msr-cards__card--default col">
<div class="card has-spectrum-border-top__hover material-card h-100 p-0">
<div class="embed-responsive embed-responsive-16by9">
<img loading="lazy" decoding="async" width="380" height="214" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1-640x360.jpg" class="card-img embed-responsive-item img-object-cover" alt="Microsoft Research Forum S2E4 | Amit Sharma | Introducing Interwhen: Steering reasoning agents with real-time verification" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1-640x360.jpg 640w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1-300x169.jpg 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1-1024x576.jpg 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1-768x432.jpg 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1-1066x600.jpg 1066w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1-655x368.jpg 655w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1-240x135.jpg 240w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1-960x540.jpg 960w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1.jpg 1280w" sizes="auto, (max-width: 380px) 100vw, 380px" /> </div>
<div class="card-body p-4 pt-3">
<h3 class="h5">
<a href="https://www.microsoft.com/en-us/research/video/introducing-interwhen-steering-reasoning-agents-with-real-time-verification/" data-bi-cN="Introducing Interwhen: Steering reasoning agents with real-time verification" class="text-decoration-none text-body"><span>Introducing Interwhen: Steering reasoning agents with real-time verification</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span></a> </h3>
<div class="card__description card__citation small">
</div>
</div>
</div>
</li>
<li class="my-0 msr-cards__card msr-cards__card--default col">
<div class="card has-spectrum-border-top__hover material-card h-100 p-0">
<div class="embed-responsive embed-responsive-16by9">
<img loading="lazy" decoding="async" width="380" height="214" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1-640x360.jpg" class="card-img embed-responsive-item img-object-cover" alt="Microsoft Research Forum S2E4 | Carles Domingo-Enrich | New fine-tuning of language models: Match meaning, not tokens" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1-640x360.jpg 640w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1-300x169.jpg 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1-1024x576.jpg 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1-768x432.jpg 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1-1066x600.jpg 1066w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1-655x368.jpg 655w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1-240x135.jpg 240w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1-960x540.jpg 960w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1.jpg 1280w" sizes="auto, (max-width: 380px) 100vw, 380px" /> </div>
<div class="card-body p-4 pt-3">
<h3 class="h5">
<a href="https://www.microsoft.com/en-us/research/video/new-fine-tuning-of-language-models-match-meaning-not-tokens/" data-bi-cN="New fine-tuning of language models: Match meaning, not tokens" class="text-decoration-none text-body"><span>New fine-tuning of language models: Match meaning, not tokens</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span></a> </h3>
<div class="card__description card__citation small">
</div>
</div>
</div>
</li>
<li class="my-0 msr-cards__card msr-cards__card--default col">
<div class="card has-spectrum-border-top__hover material-card h-100 p-0">
<div class="embed-responsive embed-responsive-16by9">
<img loading="lazy" decoding="async" width="380" height="214" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1-640x360.jpg" class="card-img embed-responsive-item img-object-cover" alt="Microsoft Research Forum S2E4 | David Rothschild | Guiding the AI disruption to the Good Place" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1-640x360.jpg 640w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1-300x169.jpg 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1-1024x576.jpg 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1-768x432.jpg 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1-1066x600.jpg 1066w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1-655x368.jpg 655w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1-240x135.jpg 240w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1-960x540.jpg 960w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1.jpg 1280w" sizes="auto, (max-width: 380px) 100vw, 380px" /> </div>
<div class="card-body p-4 pt-3">
<h3 class="h5">
<a href="https://www.microsoft.com/en-us/research/video/guiding-the-ai-disruption-to-the-good-place/" data-bi-cN="Guiding the AI disruption to the Good Place" class="text-decoration-none text-body"><span>Guiding the AI disruption to the Good Place</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span></a> </h3>
<div class="card__description card__citation small">
</div>
</div>
</div>
</li>
</ul>
</div>
</div>
</div>
<div class="wp-block-buttons is-content-justification-center is-content-justification-center is-layout-flex wp-container-core-buttons-is-layout-fe48e5de wp-block-buttons-is-layout-flex">
<div class="wp-block-button is-style-fill"><a data-bi-type="button" class="wp-block-button__link wp-element-button" href="https://aka.ms/researchforum-sessions">All Research Forum sessions</a></div>
</div> </div>
</div>
<img loading="lazy" decoding="async" width="1600" height="900" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_CatchUpBG_1920x1920-1920x1080.jpg" class="wp-block-msr-immersive-section__background-image" alt="marbled pastel background" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_CatchUpBG_1920x1920-1920x1080.jpg 1920w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_CatchUpBG_1920x1920-1066x600.jpg 1066w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_CatchUpBG_1920x1920-655x368.jpg 655w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_CatchUpBG_1920x1920-640x360.jpg 640w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_CatchUpBG_1920x1920-960x540.jpg 960w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_CatchUpBG_1920x1920-1280x720.jpg 1280w" sizes="auto, (max-width: 1600px) 100vw, 1600px" /></div>
<div style="padding-bottom:64px;padding-top:64px" class="wp-block-msr-immersive-section alignfull row">
<div class="container">
<div class="wp-block-msr-immersive-section__wrapper">
<h2 class="wp-block-heading has-text-align-center" id="explore-more-about-microsoft-research">Explore more about Microsoft Research</h2>
<div class="wp-block-msr-cards msr-cards msr-cards--default mt-4 has-text-align-left has-cards" data-bi-aN="Cards">
<div class="msr-cards__inner">
<div class="mt-4">
<ul class="my-0 list-unstyled row row-cols-1 row-cols-sm-2 row-cols-lg-4">
<li class="my-0 msr-cards__card msr-cards__card--default col">
<div class="card has-spectrum-border-top__hover material-card h-100 p-0">
<div class="embed-responsive embed-responsive-16by9">
<img loading="lazy" decoding="async" width="380" height="214" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/03/CHI_header_1920x720-640x360.png" class="card-img embed-responsive-item img-object-cover" alt="background pattern" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/03/CHI_header_1920x720-640x360.png 640w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/03/CHI_header_1920x720-1066x600.png 1066w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/03/CHI_header_1920x720-655x368.png 655w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/03/CHI_header_1920x720-960x540.png 960w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/03/CHI_header_1920x720-1280x720.png 1280w" sizes="auto, (max-width: 380px) 100vw, 380px" /> </div>
<div class="card-body p-4 pt-3">
<h3 class="h5">
<a href="https://www.microsoft.com/en-us/research/event/chi-2026/" data-bi-cN="Microsoft at CHI 2026" class="text-decoration-none text-body"><span>Microsoft at CHI 2026</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span></a> </h3>
<div class="card__description card__citation small">
<p>Explore our contributions.</p>
</div>
</div>
</div>
</li>
<li class="my-0 msr-cards__card msr-cards__card--default col">
<div class="card has-spectrum-border-top__hover material-card h-100 p-0">
<div class="embed-responsive embed-responsive-16by9">
<img loading="lazy" decoding="async" width="380" height="214" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-Careers_16-9-640x360.jpg" class="card-img embed-responsive-item img-object-cover" alt="three people talking in a relaxed setting" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-Careers_16-9-640x360.jpg 640w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-Careers_16-9-300x169.jpg 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-Careers_16-9-1024x576.jpg 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-Careers_16-9-768x432.jpg 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-Careers_16-9-1066x600.jpg 1066w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-Careers_16-9-655x368.jpg 655w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-Careers_16-9-240x135.jpg 240w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-Careers_16-9-960x540.jpg 960w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-Careers_16-9.jpg 1200w" sizes="auto, (max-width: 380px) 100vw, 380px" /> </div>
<div class="card-body p-4 pt-3">
<h3 class="h5">
<a href="https://www.microsoft.com/en-us/research/careers/" data-bi-cN="Explore careers in research" class="text-decoration-none text-body"><span>Explore careers in research</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span></a> </h3>
<div class="card__description card__citation small">
<p>Join a brilliant team of researchers and engineers working to solve technology’s most exciting challenges.</p>
</div>
</div>
</div>
</li>
<li class="my-0 msr-cards__card msr-cards__card--default col">
<div class="card has-spectrum-border-top__hover material-card h-100 p-0">
<div class="embed-responsive embed-responsive-16by9">
<img loading="lazy" decoding="async" width="380" height="214" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-MSR-Blog_16-9-640x360.png" class="card-img embed-responsive-item img-object-cover" alt="background pattern with blue and purple layers" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-MSR-Blog_16-9-640x360.png 640w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-MSR-Blog_16-9-300x169.png 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-MSR-Blog_16-9-1024x576.png 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-MSR-Blog_16-9-768x432.png 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-MSR-Blog_16-9-1066x600.png 1066w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-MSR-Blog_16-9-655x368.png 655w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-MSR-Blog_16-9-240x135.png 240w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-MSR-Blog_16-9-960x540.png 960w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/Forum-explore-MSR-Blog_16-9.png 1200w" sizes="auto, (max-width: 380px) 100vw, 380px" /> </div>
<div class="card-body p-4 pt-3">
<h3 class="h5">
<a href="https://www.microsoft.com/en-us/research/blog/" data-bi-cN="Microsoft Research Blog" class="text-decoration-none text-body"><span>Microsoft Research Blog</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span></a> </h3>
<div class="card__description card__citation small">
<p>The latest news and insights from Microsoft Research, covering topics such as AI, data science, machine learning, human-computer interaction, and more.</p>
</div>
</div>
</div>
</li>
<li class="my-0 msr-cards__card msr-cards__card--default col">
<div class="card has-spectrum-border-top__hover material-card h-100 p-0">
<div class="embed-responsive embed-responsive-16by9">
<img loading="lazy" decoding="async" width="380" height="214" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/03/SiteCard_ICLR26_1400x788-640x360.png" class="card-img embed-responsive-item img-object-cover" alt="Microsoft at ICLR logo" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2026/03/SiteCard_ICLR26_1400x788-640x360.png 640w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/03/SiteCard_ICLR26_1400x788-1066x600.png 1066w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/03/SiteCard_ICLR26_1400x788-655x368.png 655w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/03/SiteCard_ICLR26_1400x788-960x540.png 960w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/03/SiteCard_ICLR26_1400x788-1280x720.png 1280w" sizes="auto, (max-width: 380px) 100vw, 380px" /> </div>
<div class="card-body p-4 pt-3">
<h3 class="h5">
<a href="https://www.microsoft.com/en-us/research/event/iclr-2026/" data-bi-cN="Microsoft at ICLR 2026" class="text-decoration-none text-body"><span>Microsoft at ICLR 2026</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span></a> </h3>
<div class="card__description card__citation small">
<p>Explore our contributions.</p>
</div>
</div>
</div>
</li>
</ul>
</div>
</div>
</div> </div>
</div>
<img loading="lazy" decoding="async" width="1600" height="600" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/06/forum-background_1600x600_2.png" class="wp-block-msr-immersive-section__background-image" alt="background pattern" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/06/forum-background_1600x600_2.png 1600w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/06/forum-background_1600x600_2-300x113.png 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/06/forum-background_1600x600_2-1024x384.png 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/06/forum-background_1600x600_2-768x288.png 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/06/forum-background_1600x600_2-1536x576.png 1536w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/06/forum-background_1600x600_2-240x90.png 240w" sizes="auto, (max-width: 1600px) 100vw, 1600px" /></div>
<div style="padding-bottom:32px;padding-top:32px" class="wp-block-msr-immersive-section alignfull row">
<div class="container">
<div class="wp-block-msr-immersive-section__wrapper col-lg-11 col-xl-9 px-0 m-auto">
<h2 class="wp-block-heading has-text-align-center has-white-color has-text-color has-link-color wp-elements-4789963789ee9fcc3e241e4dfbbf2bfc" id="microsoft-research-forum">Microsoft Research Forum</h2>
<div class="wp-block-buttons is-content-justification-center is-content-justification-center is-layout-flex wp-container-core-buttons-is-layout-fe48e5de wp-block-buttons-is-layout-flex">
<div class="wp-block-button"><a data-bi-type="button" class="wp-block-button__link wp-element-button" href="http://aka.ms/researchforum-register" target="_blank" rel="noreferrer noopener">Register for the series</a></div>
<div class="wp-block-button"><a data-bi-type="button" class="wp-block-button__link wp-element-button" href="https://www.microsoft.com/en-us/research/event/microsoft-research-forum/event-code-of-conduct/">Event Code of Conduct</a></div>
</div> </div>
</div>
<img loading="lazy" decoding="async" width="1600" height="600" src="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_LowerPromoBanner_1920x720.jpg" class="wp-block-msr-immersive-section__background-image" alt="dark blue watercolor background" srcset="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_LowerPromoBanner_1920x720.jpg 1920w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_LowerPromoBanner_1920x720-300x113.jpg 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_LowerPromoBanner_1920x720-1024x384.jpg 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_LowerPromoBanner_1920x720-768x288.jpg 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_LowerPromoBanner_1920x720-1536x576.jpg 1536w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_LowerPromoBanner_1920x720-1600x600.jpg 1600w, https://www.microsoft.com/en-us/research/wp-content/uploads/2025/04/MRF_MCR_LowerPromoBanner_1920x720-240x90.jpg 240w" sizes="auto, (max-width: 1600px) 100vw, 1600px" /></div>
</div>
</div>
</section>
</div>
<span id="label-external-link" class="sr-only" aria-hidden="true">Opens in a new tab</span> </main>
<div ms.pgarea="social" data-moray>
<section class="msr-social msr-social--footer py-3" role="region" aria-label="Social media links"
data-bi-aN="SocialMediaLinks">
<div class="container">
<div class="row">
<div class="col-12 col-md-6 msr-social-col msr-social-col--follow">
<div class="d-flex flex-row flex-wrap align-items-center">
<p class="mr-2 mb-0" id="msr-follow-us-footer">
Follow us: </p>
<ul class="list-unstyled d-inline-flex flex-row-auto gap-2 align-items-center mb-0" aria-labelledby="msr-follow-us-footer">
<li class="mr-0 mb-0 p-1">
<a
href="https://x.com/intent/follow?original_referrer=https%3A%2F%2Fwww.microsoft.com%2Fen-us%2Fresearch%2Fevent%2Fmicrosoft-research-forum%2F&screen_name=MSFTResearch"
data-bi-slot="0"
data-bi-cN="Follow on X"
data-bi-type="social-link"
data-bi-tN="social-follow"
data-bi-bhvr="126"
target="_blank"
rel="noopener noreferrer"
class="d-block"
>
<span class="sr-only">Follow on X</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M18.42,14.009L27.891,3h-2.244l-8.224,9.559L10.855,3H3.28l9.932,14.455L3.28,29h2.244l8.684-10.095,6.936,10.095h7.576l-10.301-14.991h0Zm-3.074,3.573l-1.006-1.439L6.333,4.69h3.447l6.462,9.243,1.006,1.439,8.4,12.015h-3.447l-6.854-9.804h0Z" /></svg>
</a>
</li>
<li class="mr-0 mb-0 p-1">
<a
href="https://www.facebook.com/microsoftresearch/"
data-bi-slot="1"
data-bi-cN="Like on Facebook"
data-bi-type="social-link"
data-bi-tN="social-follow"
data-bi-bhvr="126"
target="_blank"
rel="noopener noreferrer"
class="d-block"
>
<span class="sr-only">Like on Facebook</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M16,2c-7.732,0-14,6.268-14,14,0,6.566,4.52,12.075,10.618,13.588v-9.31h-2.887v-4.278h2.887v-1.843c0-4.765,2.156-6.974,6.835-6.974,.887,0,2.417,.174,3.043,.348v3.878c-.33-.035-.904-.052-1.617-.052-2.296,0-3.183,.87-3.183,3.13v1.513h4.573l-.786,4.278h-3.787v9.619c6.932-.837,12.304-6.74,12.304-13.897,0-7.732-6.268-14-14-14Z" /></svg>
</a>
</li>
<li class="mb-0 p-1">
<a
href="https://www.linkedin.com/showcase/microsoftresearch/"
data-bi-slot="5"
data-bi-cN="Follow on LinkedIn"
data-bi-type="social-link"
data-bi-tN="social-follow"
data-bi-bhvr="126"
target="_blank"
rel="noopener noreferrer"
class="d-block"
>
<span class="sr-only">Follow on LinkedIn</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M26.111,3H5.889c-1.595,0-2.889,1.293-2.889,2.889V26.111c0,1.595,1.293,2.889,2.889,2.889H26.111c1.595,0,2.889-1.293,2.889-2.889V5.889c0-1.595-1.293-2.889-2.889-2.889ZM10.861,25.389h-3.877V12.87h3.877v12.519Zm-1.957-14.158c-1.267,0-2.293-1.034-2.293-2.31s1.026-2.31,2.293-2.31,2.292,1.034,2.292,2.31-1.026,2.31-2.292,2.31Zm16.485,14.158h-3.858v-6.571c0-1.802-.685-2.809-2.111-2.809-1.551,0-2.362,1.048-2.362,2.809v6.571h-3.718V12.87h3.718v1.686s1.118-2.069,3.775-2.069,4.556,1.621,4.556,4.975v7.926Z" fill-rule="evenodd" /></svg>
</a>
</li>
<li class="mb-0 p-1">
<a
href="https://www.youtube.com/user/MicrosoftResearch"
data-bi-slot="2"
data-bi-cN="Subscribe on Youtube"
data-bi-type="social-link"
data-bi-tN="social-follow"
data-bi-bhvr="126"
target="_blank"
rel="noopener noreferrer"
class="d-block"
>
<span class="sr-only">Subscribe on Youtube</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M31.331,8.248c-.368-1.386-1.452-2.477-2.829-2.848-2.496-.673-12.502-.673-12.502-.673,0,0-10.007,0-12.502,.673-1.377,.37-2.461,1.462-2.829,2.848-.669,2.512-.669,7.752-.669,7.752,0,0,0,5.241,.669,7.752,.368,1.386,1.452,2.477,2.829,2.847,2.496,.673,12.502,.673,12.502,.673,0,0,10.007,0,12.502-.673,1.377-.37,2.461-1.462,2.829-2.847,.669-2.512,.669-7.752,.669-7.752,0,0,0-5.24-.669-7.752ZM12.727,20.758V11.242l8.364,4.758-8.364,4.758Z" fill="currentColor" /></svg>
</a>
</li>
<li class="mb-0 p-1">
<a
href="https://www.instagram.com/msft_research/"
data-bi-slot="3"
data-bi-cN="Follow on Instagram"
data-bi-type="social-link"
data-bi-tN="social-follow"
data-bi-bhvr="126"
target="_blank"
rel="noopener noreferrer"
class="d-block"
>
<span class="sr-only">Follow on Instagram</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M10.202,2.098c-1.49,.07-2.507,.308-3.396,.657-.92,.359-1.7,.84-2.477,1.619-.776,.779-1.254,1.56-1.61,2.481-.345,.891-.578,1.909-.644,3.4-.066,1.49-.08,1.97-.073,5.771s.024,4.278,.096,5.772c.071,1.489,.308,2.506,.657,3.396,.359,.92,.84,1.7,1.619,2.477,.779,.776,1.559,1.253,2.483,1.61,.89,.344,1.909,.579,3.399,.644,1.49,.065,1.97,.08,5.771,.073,3.801-.007,4.279-.024,5.773-.095s2.505-.309,3.395-.657c.92-.36,1.701-.84,2.477-1.62s1.254-1.561,1.609-2.483c.345-.89,.579-1.909,.644-3.398,.065-1.494,.081-1.971,.073-5.773s-.024-4.278-.095-5.771-.308-2.507-.657-3.397c-.36-.92-.84-1.7-1.619-2.477s-1.561-1.254-2.483-1.609c-.891-.345-1.909-.58-3.399-.644s-1.97-.081-5.772-.074-4.278,.024-5.771,.096m.164,25.309c-1.365-.059-2.106-.286-2.6-.476-.654-.252-1.12-.557-1.612-1.044s-.795-.955-1.05-1.608c-.192-.494-.423-1.234-.487-2.599-.069-1.475-.084-1.918-.092-5.656s.006-4.18,.071-5.656c.058-1.364,.286-2.106,.476-2.6,.252-.655,.556-1.12,1.044-1.612s.955-.795,1.608-1.05c.493-.193,1.234-.422,2.598-.487,1.476-.07,1.919-.084,5.656-.092,3.737-.008,4.181,.006,5.658,.071,1.364,.059,2.106,.285,2.599,.476,.654,.252,1.12,.555,1.612,1.044s.795,.954,1.051,1.609c.193,.492,.422,1.232,.486,2.597,.07,1.476,.086,1.919,.093,5.656,.007,3.737-.006,4.181-.071,5.656-.06,1.365-.286,2.106-.476,2.601-.252,.654-.556,1.12-1.045,1.612s-.955,.795-1.608,1.05c-.493,.192-1.234,.422-2.597,.487-1.476,.069-1.919,.084-5.657,.092s-4.18-.007-5.656-.071M21.779,8.517c.002,.928,.755,1.679,1.683,1.677s1.679-.755,1.677-1.683c-.002-.928-.755-1.679-1.683-1.677,0,0,0,0,0,0-.928,.002-1.678,.755-1.677,1.683m-12.967,7.496c.008,3.97,3.232,7.182,7.202,7.174s7.183-3.232,7.176-7.202c-.008-3.97-3.233-7.183-7.203-7.175s-7.182,3.233-7.174,7.203m2.522-.005c-.005-2.577,2.08-4.671,4.658-4.676,2.577-.005,4.671,2.08,4.676,4.658,.005,2.577-2.08,4.671-4.658,4.676-2.577,.005-4.671-2.079-4.676-4.656h0" /></svg>
</a>
</li>
<li class="mb-0 p-1">
<a
href="https://www.microsoft.com/en-us/research/feed/"
data-bi-slot="4"
data-bi-cN="Subscribe to our RSS feed"
data-bi-type="social-link"
data-bi-tN="social-follow"
data-bi-bhvr="126"
target="_blank"
rel="noopener noreferrer"
class="d-block"
>
<span class="sr-only">Subscribe to our RSS feed</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><circle cx="6.566" cy="25.434" r="3.566" fill="currentColor" /><path d="M20.234,29h-5.051c0-6.728-5.454-12.183-12.183-12.183h0v-5.051c9.518,0,17.234,7.716,17.234,17.234Z" fill="currentColor" /><path d="M23.8,29c0-11.488-9.312-20.8-20.8-20.8V3c14.359,0,26,11.641,26,26h-5.2Z" fill="currentColor" /></svg>
</a>
</li>
</ul>
</div>
</div><!--/.col-->
<div class="col-12 col-md-6 msr-social-col msr-social-col--share mt-3 mt-md-0">
<div class="d-flex flex-row flex-wrap align-items-center">
<p class="mr-3 mb-0" id="msr-share-footer">
Share this page: </p>
<ul class="list-unstyled d-flex gap-2 align-items-center mb-0" aria-labelledby="msr-share-footer">
<li class="mr-0 mb-0 p-1">
<a
href="https://x.com/intent/tweet?text=Microsoft%20Research%20Forum&url=https%3A%2F%2Fwww.microsoft.com%2Fen-us%2Fresearch%2Fevent%2Fmicrosoft-research-forum%2F"
data-bi-slot="5"
data-bi-cN="Share on X"
data-bi-type="social-link"
data-bi-tN="social-share"
data-bi-bhvr="120"
target="_blank"
rel="noopener noreferrer"
class="d-block"
>
<span class="sr-only">Share on X</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M18.42,14.009L27.891,3h-2.244l-8.224,9.559L10.855,3H3.28l9.932,14.455L3.28,29h2.244l8.684-10.095,6.936,10.095h7.576l-10.301-14.991h0Zm-3.074,3.573l-1.006-1.439L6.333,4.69h3.447l6.462,9.243,1.006,1.439,8.4,12.015h-3.447l-6.854-9.804h0Z" /></svg>
</a>
</li>
<li class="mr-0 mb-0 p-1">
<a
href="https://www.facebook.com/sharer/sharer.php?u=https%3A%2F%2Fwww.microsoft.com%2Fen-us%2Fresearch%2Fevent%2Fmicrosoft-research-forum%2F"
data-bi-slot="6"
data-bi-cN="Share on Facebook"
data-bi-type="social-link"
data-bi-tN="social-share"
data-bi-bhvr="120"
target="_blank"
rel="noopener noreferrer"
class="d-block">
<span class="sr-only">Share on Facebook</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M16,2c-7.732,0-14,6.268-14,14,0,6.566,4.52,12.075,10.618,13.588v-9.31h-2.887v-4.278h2.887v-1.843c0-4.765,2.156-6.974,6.835-6.974,.887,0,2.417,.174,3.043,.348v3.878c-.33-.035-.904-.052-1.617-.052-2.296,0-3.183,.87-3.183,3.13v1.513h4.573l-.786,4.278h-3.787v9.619c6.932-.837,12.304-6.74,12.304-13.897,0-7.732-6.268-14-14-14Z" /></svg>
</a>
</li>
<li class="mb-0 p-1">
<a
href="
https://www.linkedin.com/shareArticle?mini=true&url=https%3A%2F%2Fwww.microsoft.com%2Fen-us%2Fresearch%2Fevent%2Fmicrosoft-research-forum%2F&title=Microsoft%20Research%20Forum&summary=Microsoft%20Research%20Forum&source=Microsoft%20Research "
data-bi-slot="7"
data-bi-cN="Share on LinkedIn"
data-bi-type="social-link"
data-bi-tN="social-share"
data-bi-bhvr="120"
target="_blank"
rel="noopener noreferrer"
class="d-block"
>
<span class="sr-only">Share on LinkedIn</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M26.111,3H5.889c-1.595,0-2.889,1.293-2.889,2.889V26.111c0,1.595,1.293,2.889,2.889,2.889H26.111c1.595,0,2.889-1.293,2.889-2.889V5.889c0-1.595-1.293-2.889-2.889-2.889ZM10.861,25.389h-3.877V12.87h3.877v12.519Zm-1.957-14.158c-1.267,0-2.293-1.034-2.293-2.31s1.026-2.31,2.293-2.31,2.292,1.034,2.292,2.31-1.026,2.31-2.292,2.31Zm16.485,14.158h-3.858v-6.571c0-1.802-.685-2.809-2.111-2.809-1.551,0-2.362,1.048-2.362,2.809v6.571h-3.718V12.87h3.718v1.686s1.118-2.069,3.775-2.069,4.556,1.621,4.556,4.975v7.926Z" fill-rule="evenodd" /></svg>
</a>
</li>
<li class="mb-0 p-1">
<a href="
http://www.reddit.com/submit?title=Microsoft%20Research%20Forum&url=https%3A%2F%2Fwww.microsoft.com%2Fen-us%2Fresearch%2Fevent%2Fmicrosoft-research-forum%2F "
data-bi-slot="8"
data-bi-cN="Share on Reddit"
data-bi-type="social-link"
data-bi-tN="social-share"
data-bi-bhvr="120"
target="_blank"
rel="noopener noreferrer"
class="d-block"
>
<span class="sr-only">Share on Reddit</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M27.332,10.323c-1.07,0-2.055,.361-2.842,.967-2.143-1.326-4.848-2.16-7.807-2.271v-.013c0-1.983,1.474-3.629,3.386-3.9v-.003c.347,1.47,1.666,2.564,3.242,2.564,1.84,0,3.331-1.491,3.331-3.331s-1.491-3.331-3.331-3.331c-1.609,0-2.95,1.14-3.262,2.657-2.694,.289-4.798,2.574-4.798,5.343v.017c-2.93,.123-5.605,.957-7.729,2.274-.789-.611-1.779-.974-2.853-.974-2.578,0-4.668,2.09-4.668,4.668,0,1.871,1.099,3.483,2.688,4.228,.155,5.419,6.06,9.778,13.323,9.778s13.176-4.364,13.323-9.787c1.576-.75,2.666-2.357,2.666-4.217,0-2.578-2.09-4.668-4.668-4.668ZM7.334,17.952c.078-1.693,1.203-2.992,2.51-2.992s2.307,1.373,2.229,3.066c-.078,1.693-1.054,2.308-2.363,2.308s-2.453-.689-2.375-2.382Zm13.596,4.424c-.804,1.922-2.703,3.273-4.919,3.273s-4.114-1.351-4.919-3.273c-.095-.228,.061-.483,.306-.508,1.437-.145,2.991-.225,4.613-.225s3.175,.08,4.613,.225c.245,.025,.401,.28,.306,.508Zm1.384-2.043c-1.307,0-2.285-.614-2.363-2.308-.078-1.693,.92-3.066,2.229-3.066s2.433,1.299,2.51,2.992c.078,1.693-1.068,2.382-2.375,2.382Z" /></svg>
</a>
</li>
</ul>
</div>
</div><!--/.col-->
</div><!--/.row-->
</div><!--/.container-->
</section><!--/.ms-social-->
</div>
<div id="playerModal" class="mfp-hide">
<div id="player"></div>
</div>
<div id="mq"></div>
<uhf-footer locale="en-us" partnerId="MSRESEARCH" footerId="global-default-footer" theme="light">
<script id="uhf-footer-ccpa">
(function () {
function checkThirdPartyAdsOptOutCookie() {
try {
var match = document.cookie.match('(^|;)\\s*3PAdsOptOut\\s*=\\s*([^;]+)');
return (match ? match[2] : '') !== '1';
} catch (e) {
return true;
}
}
var globalPrivacyControlEnabled = navigator.globalPrivacyControl;
window.GPC_DataSharingOptIn = globalPrivacyControlEnabled ? false : checkThirdPartyAdsOptOutCookie();
if (typeof window.onGPCLoaded === 'function') {
window.onGPCLoaded();
}
})();
</script>
<uhf-footer-nav slot="uhf-footer-nav">
<div class="uhf-footer-nav-row">
<uhf-footer-nav-group class="uhf-footer-nav-group" heading="What's new" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_What's new_nav", "ecn": "FooterNav_Whats_New_nav", "ehn": "FooterNav"}">
<a class="uhf-footer-link" href="https://www.microsoft.com/surface/devices/surface-pro" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_What's new_Surface Pro_nav", "ecn": "FooterNav_Whats_New_Footer_WhatsNew_NewSurfacePro_nav", "ehn": "FooterNav"}">Surface Pro</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/surface/devices/surface-laptop" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_What's new_Surface Laptop_nav", "ecn": "FooterNav_Whats_New_Footer_WhatsNew_SurfaceLaptop_nav", "ehn": "FooterNav"}">Surface Laptop</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/surface/devices/surface-laptop-ultra?icid=DSM_Footer_WhatsNew_SurfaceLaptopUltra" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_What's new_Surface Laptop Ultra_nav", "ecn": "FooterNav_Whats_New_Footer_WhatsNew_Surface Laptop Ultra_nav", "ehn": "FooterNav"}">Surface Laptop Ultra</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/surface/devices/surface-rtx-spark-dev-box?icid=DSM_Footer_WhatsNew_SurfaceRTXSparkDevBox" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_What's new_Surface RTX Spark Dev Box_nav", "ecn": "FooterNav_Whats_New_footer_whatsnew_SurfaceRTXSparksDevBox_nav_nav", "ehn": "FooterNav"}">Surface RTX Spark Dev Box</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/microsoft-copilot/organizations?icid=DSM_Footer_CopilotOrganizations" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_What's new_Copilot for organizations_nav", "ecn": "FooterNav_Whats_New_Footer_WhatsNew_CopilotMicrosoft_nav", "ehn": "FooterNav"}">Copilot for organizations</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/microsoft-copilot/for-individuals?form=MY02PT&OCID=GE_web_Copilot_Free_868g3t5nj" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_What's new_Copilot for personal use_nav", "ecn": "FooterNav_Whats_New_Footer_WhatsNew_CopilotPersonal_nav", "ehn": "FooterNav"}">Copilot for personal use</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/microsoft-products-and-apps" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_What's new_Explore Microsoft products_nav", "ecn": "FooterNav_Whats_New_Footer_WhatsNew_ExploreMicrosoftProducts_nav", "ehn": "FooterNav"}">Explore Microsoft products</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/windows/apps-for-windows?icid=DSM_Footer_WhatsNew_Windows11apps" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_What's new_Windows 11 apps_nav", "ecn": "FooterNav_Whats_New_Footer_WhatsNew_Windows_11_apps_nav", "ehn": "FooterNav"}">Windows 11 apps</a>
</uhf-footer-nav-group>
<uhf-footer-nav-group class="uhf-footer-nav-group" heading="Microsoft Store" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Microsoft Store_nav", "ecn": "FooterNav_Store_and_Support_nav", "ehn": "FooterNav"}">
<a class="uhf-footer-link" href="https://account.microsoft.com/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Microsoft Store_Account profile_nav", "ecn": "FooterNav_Store_and_Support_Footer_StoreandSupport_AccountProfile_nav", "ehn": "FooterNav"}">Account profile</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/download" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Microsoft Store_Download Center_nav", "ecn": "FooterNav_Store_and_Support_Footer_StoreandSupport_DownloadCenter_nav", "ehn": "FooterNav"}">Download Center</a>
<a class="uhf-footer-link" href="https://go.microsoft.com/fwlink/?linkid=2139749" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Microsoft Store_Microsoft Store support_nav", "ecn": "FooterNav_Store_and_Support_Footer_StoreandSupport_SalesAndSupport_nav", "ehn": "FooterNav"}">Microsoft Store support</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/store/b/returns" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Microsoft Store_Returns_nav", "ecn": "FooterNav_Store_and_Support_Footer_StoreandSupport_Returns_nav", "ehn": "FooterNav"}">Returns</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/store/b/order-tracking" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Microsoft Store_Order tracking_nav", "ecn": "FooterNav_Store_and_Support_Footer_StoreandSupport_OrderTracking_nav", "ehn": "FooterNav"}">Order tracking</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/store/b/certified-refurbished-products" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Microsoft Store_Certified Refurbished_nav", "ecn": "FooterNav_Store_and_Support_Footer_StoreandSupport_StoreLocations_nav", "ehn": "FooterNav"}">Certified Refurbished</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/store/b/why-microsoft-store?icid=footer_why-msft-store_7102020" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Microsoft Store_Microsoft Store Promise_nav", "ecn": "FooterNav_Store_and_Support_Footer_StoreandSupport_MicrosoftPromise_nav", "ehn": "FooterNav"}">Microsoft Store Promise</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/store/b/payment-financing-options?icid=footer_financing_vcc" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Microsoft Store_Flexible Payments_nav", "ecn": "FooterNav_Store_and_Support_Footer_StoreandSupport_Financing_nav", "ehn": "FooterNav"}">Flexible Payments</a>
</uhf-footer-nav-group>
<uhf-footer-nav-group class="uhf-footer-nav-group" heading="Education" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Education_nav", "ecn": "FooterNav_Education_nav", "ehn": "FooterNav"}">
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/education" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Education_Microsoft in education_nav", "ecn": "FooterNav_Education_Footer_Education_MicrosoftInEducation_nav", "ehn": "FooterNav"}">Microsoft in education</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/education/devices/overview" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Education_Devices for education_nav", "ecn": "FooterNav_Education_Footer_Education_DevicesforEducation_nav", "ehn": "FooterNav"}">Devices for education</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/education/products/teams" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Education_Microsoft Teams for Education_nav", "ecn": "FooterNav_Education_Footer_Education_MicrosoftTeamsforEducation_nav", "ehn": "FooterNav"}">Microsoft Teams for Education</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/education/products/microsoft-365" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Education_Microsoft 365 Education_nav", "ecn": "FooterNav_Education_Footer_Education_Microsoft365Education_nav", "ehn": "FooterNav"}">Microsoft 365 Education</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/education/how-to-buy" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Education_How to buy for your school_nav", "ecn": "FooterNav_Education_Footer_Howtobuyforyourschool_nav", "ehn": "FooterNav"}">How to buy for your school</a>
<a class="uhf-footer-link" href="https://education.microsoft.com/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Education_Educator training and development_nav", "ecn": "FooterNav_Education_Footer_Education_EducatorTrainingDevelopment_nav", "ehn": "FooterNav"}">Educator training and development</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/store/b/education" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Education_Deals for students and parents_nav", "ecn": "FooterNav_Education_Footer_Education_DealsForStudentsandParents_nav", "ehn": "FooterNav"}">Deals for students and parents</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/education/ai-in-education" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Education_AI for education_nav", "ecn": "FooterNav_Education_Footer_Education_Azureforstudents_nav", "ehn": "FooterNav"}">AI for education</a>
</uhf-footer-nav-group>
</div>
<div class="uhf-footer-nav-row">
<uhf-footer-nav-group class="uhf-footer-nav-group" heading="Business" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Business_nav", "ecn": "FooterNav_Business_nav", "ehn": "FooterNav"}">
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/ai?icid=DSM_Footer_AI" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Business_Microsoft AI_nav", "ecn": "FooterNav_Business_Footer_Business_AI_nav", "ehn": "FooterNav"}">Microsoft AI</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/security" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Business_Microsoft Security_nav", "ecn": "FooterNav_Business_Footer_Business_Microsoft Security_nav", "ehn": "FooterNav"}">Microsoft Security</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/dynamics-365" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Business_Dynamics 365_nav", "ecn": "FooterNav_Business_Footer_Business_MicrosoftDynamics365_nav", "ehn": "FooterNav"}">Dynamics 365</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/microsoft-365/business" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Business_Microsoft 365_nav", "ecn": "FooterNav_Business_Footer_Business_M365_nav", "ehn": "FooterNav"}">Microsoft 365</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/power-platform" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Business_Microsoft Power Platform_nav", "ecn": "FooterNav_Business_Footer_DeveloperAndIT_Power Platform_nav", "ehn": "FooterNav"}">Microsoft Power Platform</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/microsoft-teams/group-chat-software" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Business_Microsoft Teams_nav", "ecn": "FooterNav_Business_Footer_Business_Microsoft365_nav", "ehn": "FooterNav"}">Microsoft Teams</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/microsoft-365-copilot?icid=DSM_Footer_Microsoft365Copilot" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Business_Microsoft 365 Copilot_nav", "ecn": "FooterNav_Business_Footer_CopilotMicrosoft365 _nav", "ehn": "FooterNav"}">Microsoft 365 Copilot</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/store/b/business?icid=CNavBusinessStore" id="Small-Business" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Business_Small Business_nav", "ecn": "FooterNav_Business_Footer_Business-SmallBusiness_nav", "ehn": "FooterNav"}">Small Business</a>
</uhf-footer-nav-group>
<uhf-footer-nav-group class="uhf-footer-nav-group" heading="Developer & IT" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Developer & IT_nav", "ecn": "FooterNav_Developer_nav", "ehn": "FooterNav"}">
<a class="uhf-footer-link" href="https://azure.microsoft.com/en-us/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Developer & IT_Azure_nav", "ecn": "FooterNav_Developer_Footer_DeveloperAndIT_MicrosoftAzure_nav", "ehn": "FooterNav"}">Azure</a>
<a class="uhf-footer-link" href="https://developer.microsoft.com/en-us/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Developer & IT_Microsoft Developer_nav", "ecn": "FooterNav_Developer_Footer_DeveloperAndIT_MicrosoftDeveloper_nav", "ehn": "FooterNav"}">Microsoft Developer</a>
<a class="uhf-footer-link" href="https://learn.microsoft.com/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Developer & IT_Microsoft Learn_nav", "ecn": "FooterNav_Developer_Footer_DeveloperAndIT_MicrosoftLearn_nav", "ehn": "FooterNav"}">Microsoft Learn</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/software-development-companies/offers-benefits/isv-success?icid=DSM_Footer_SupportAIMarketplace&ocid=cmm3atxvn98" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Developer & IT_Support for AI marketplace apps_nav", "ecn": "FooterNav_Developer_Footer_DeveloperAndIT_SupportForAIMarketplaceApps_nav", "ehn": "FooterNav"}">Support for AI marketplace apps</a>
<a class="uhf-footer-link" href="https://techcommunity.microsoft.com/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Developer & IT_Microsoft Tech Community_nav", "ecn": "FooterNav_Developer_Footer_DeveloperAndIT_MicrosoftTechCommunity_nav", "ehn": "FooterNav"}">Microsoft Tech Community</a>
<a class="uhf-footer-link" href="https://marketplace.microsoft.com?icid=DSM_Footer_Marketplace&ocid=cmm3atxvn98" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Developer & IT_Microsoft Marketplace_nav", "ecn": "FooterNav_Developer_Footer_DeveloperAndIT_Marketplace_nav", "ehn": "FooterNav"}">Microsoft Marketplace</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/software-development-companies?icid=DSM_Footer_SoftwareCompanies&ocid=cmm3atxvn98" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Developer & IT_Software companies_nav", "ecn": "FooterNav_Developer_Software companies_nav", "ehn": "FooterNav"}">Software companies</a>
<a class="uhf-footer-link" href="https://visualstudio.microsoft.com/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Developer & IT_Visual Studio_nav", "ecn": "FooterNav_Developer_Footer_DeveloperAndIT_MicrosoftVisualStudio_nav", "ehn": "FooterNav"}">Visual Studio</a>
</uhf-footer-nav-group>
<uhf-footer-nav-group class="uhf-footer-nav-group" heading="Company" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Company_nav", "ecn": "FooterNav_Company_nav", "ehn": "FooterNav"}">
<a class="uhf-footer-link" href="https://careers.microsoft.com/" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Company_Careers_nav", "ecn": "FooterNav_Company_Footer_Company_Careers_nav", "ehn": "FooterNav"}">Careers</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/about" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Company_About Microsoft_nav", "ecn": "FooterNav_Company_Footer_Company_AboutMicrosoft_nav", "ehn": "FooterNav"}">About Microsoft</a>
<a class="uhf-footer-link" href="https://news.microsoft.com/source/?icid=DSM_Footer_Company_CompanyNews" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Company_Company news_nav", "ecn": "FooterNav_Company_Footer_Company_CompanyNews_nav", "ehn": "FooterNav"}">Company news</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/privacy?icid=DSM_Footer_Company_Privacy" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Company_Privacy at Microsoft_nav", "ecn": "FooterNav_Company_Footer_Company_PrivacyAtMicrosoft_nav", "ehn": "FooterNav"}">Privacy at Microsoft</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/investor/default.aspx" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Company_Investors_nav", "ecn": "FooterNav_Company_Footer_Company_Investors_nav", "ehn": "FooterNav"}">Investors</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/diversity/default?icid=DSM_Footer_Company_Diversity" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Company_Diversity and inclusion_nav", "ecn": "FooterNav_Company_Footer_Company_DiversityAndInclusion_nav", "ehn": "FooterNav"}">Diversity and inclusion</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/accessibility" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Company_Accessibility_nav", "ecn": "FooterNav_Company_Footer_Company_Accessibility_nav", "ehn": "FooterNav"}">Accessibility</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/corporate-responsibility/sustainability?icid=DSM_Footer_Sustainability" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "FooterNav", "cN": "FooterNav_Company_Sustainability_nav", "ecn": "FooterNav_Company_Footer_Company_Sustainability_nav", "ehn": "FooterNav"}">Sustainability</a>
</uhf-footer-nav-group>
</div>
</uhf-footer-nav>
<div slot="uhf-footer-california-privacy-link">
<a class="uhf-footer-link" href="https://aka.ms/yourcaliforniaprivacychoices" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_California Privacy_nav", "ecn": "LegalNav_CaliforniaPrivacy_nav", "ehn": "LegalNav"}">
<svg role="img" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 30 14" xml:space="preserve" height="16" width="43">
<title>Your Privacy Choices Opt-Out Icon</title>
<path d="M7.4 12.8h6.8l3.1-11.6H7.4C4.2 1.2 1.6 3.8 1.6 7s2.6 5.8 5.8 5.8z" style="fill-rule:evenodd;clip-rule:evenodd;fill:#fff"/>
<path d="M22.6 0H7.4c-3.9 0-7 3.1-7 7s3.1 7 7 7h15.2c3.9 0 7-3.1 7-7s-3.2-7-7-7zm-21 7c0-3.2 2.6-5.8 5.8-5.8h9.9l-3.1 11.6H7.4c-3.2 0-5.8-2.6-5.8-5.8z" style="fill-rule:evenodd;clip-rule:evenodd;fill:#06f"/>
<path d="M24.6 4c.2.2.2.6 0 .8L22.5 7l2.2 2.2c.2.2.2.6 0 .8-.2.2-.6.2-.8 0l-2.2-2.2-2.2 2.2c-.2.2-.6.2-.8 0-.2-.2-.2-.6 0-.8L20.8 7l-2.2-2.2c-.2-.2-.2-.6 0-.8.2-.2.6-.2.8 0l2.2 2.2L23.8 4c.2-.2.6-.2.8 0z" style="fill:#fff"/>
<path d="M12.7 4.1c.2.2.3.6.1.8L8.6 9.8c-.1.1-.2.2-.3.2-.2.1-.5.1-.7-.1L5.4 7.7c-.2-.2-.2-.6 0-.8.2-.2.6-.2.8 0L8 8.6l3.8-4.5c.2-.2.6-.2.9 0z" style="fill:#06f"/>
</svg>
<span>Your Privacy Choices</span>
</a>
<noscript>
<a class="uhf-footer-link" href="https://aka.ms/yourcaliforniaprivacychoices" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_California Privacy_nav", "ecn": "LegalNav_CaliforniaPrivacy_nav", "ehn": "LegalNav"}">
<svg role="img" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 30 14" xml:space="preserve" height="16" width="43">
<title>Your Privacy Choices Opt-Out Icon</title>
<path d="M7.4 12.8h6.8l3.1-11.6H7.4C4.2 1.2 1.6 3.8 1.6 7s2.6 5.8 5.8 5.8z" style="fill-rule:evenodd;clip-rule:evenodd;fill:#fff"/>
<path d="M22.6 0H7.4c-3.9 0-7 3.1-7 7s3.1 7 7 7h15.2c3.9 0 7-3.1 7-7s-3.2-7-7-7zm-21 7c0-3.2 2.6-5.8 5.8-5.8h9.9l-3.1 11.6H7.4c-3.2 0-5.8-2.6-5.8-5.8z" style="fill-rule:evenodd;clip-rule:evenodd;fill:#06f"/>
<path d="M24.6 4c.2.2.2.6 0 .8L22.5 7l2.2 2.2c.2.2.2.6 0 .8-.2.2-.6.2-.8 0l-2.2-2.2-2.2 2.2c-.2.2-.6.2-.8 0-.2-.2-.2-.6 0-.8L20.8 7l-2.2-2.2c-.2-.2-.2-.6 0-.8.2-.2.6-.2.8 0l2.2 2.2L23.8 4c.2-.2.6-.2.8 0z" style="fill:#fff"/>
<path d="M12.7 4.1c.2.2.3.6.1.8L8.6 9.8c-.1.1-.2.2-.3.2-.2.1-.5.1-.7-.1L5.4 7.7c-.2-.2-.2-.6 0-.8.2-.2.6-.2.8 0L8 8.6l3.8-4.5c.2-.2.6-.2.9 0z" style="fill:#06f"/>
</svg>
<span>Your Privacy Choices</span>
</a>
</noscript>
</div>
<a slot="uhf-footer-consumer-health-privacy-link" class="uhf-footer-link" href="https://go.microsoft.com/fwlink/?linkid=2259814" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_Consumer Health Privacy_nav", "ecn": "LegalNav_ConsumerHealthPrivacy_nav", "ehn": "LegalNav"}">Consumer Health Privacy</a>
<uhf-footer-menu slot="uhf-footer-menu" data-nav-label="Microsoft corporate links">
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/sitemap1.aspx" id="uhf-Footer_Sitemap" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_Sitemap_nav", "ecn": "LegalNav_Footer_Sitemap_nav", "ehn": "LegalNav"}">Sitemap</a>
<a class="uhf-footer-link" href="https://support.microsoft.com/contactus" id="uhf-Footer_ContactUs" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_Contact Microsoft_nav", "ecn": "LegalNav_Footer_ContactUs_nav", "ehn": "LegalNav"}">Contact Microsoft</a>
<a class="uhf-footer-link" href="https://go.microsoft.com/fwlink/?LinkId=521839" id="uhf-Footer_PrivacyandCookies" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_Privacy _nav", "ecn": "LegalNav_Footer_PrivacyandCookies_nav", "ehn": "LegalNav"}">Privacy </a>
<a class="uhf-footer-link" href="#" id="uhf-Footer_ManageCookies" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_Manage cookies_nav", "ecn": "LegalNav_Footer_ManageCookies_nav", "ehn": "LegalNav"}">Manage cookies</a>
<a class="uhf-footer-link" href="https://go.microsoft.com/fwlink/?LinkID=206977" id="uhf-Footer_TermsOfUse" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_Terms of use_nav", "ecn": "LegalNav_Footer_TermsOfUse_nav", "ehn": "LegalNav"}">Terms of use</a>
<a class="uhf-footer-link" href="https://go.microsoft.com/fwlink/?linkid=2196228" id="uhf-Footer_Trademarks" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_Trademarks_nav", "ecn": "LegalNav_Footer_Trademarks_nav", "ehn": "LegalNav"}">Trademarks</a>
<a class="uhf-footer-link" href="https://go.microsoft.com/fwlink/?linkid=2196227" id="uhf-Footer_SafetyAndEco" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_Safety & eco_nav", "ecn": "LegalNav_Footer_SafetyAndEco_nav", "ehn": "LegalNav"}">Safety & eco</a>
<a class="uhf-footer-link" href="https://www.microsoft.com/en-us/legal/compliance/recycling" id="uhf-Recycling" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_Recycling_nav", "ecn": "LegalNav_Recycling_nav", "ehn": "LegalNav"}">Recycling</a>
<a class="uhf-footer-link" href="https://choice.microsoft.com" id="uhf-Footer_AboutourAds" data-m="{"compnm": "UHF", "view": "UHF", "pa": "UniversalFooter", "hn": "LegalNav", "cN": "LegalNav_About our ads_nav", "ecn": "LegalNav_Footer_AboutourAds_nav", "ehn": "LegalNav"}">About our ads</a>
</uhf-footer-menu>
</uhf-footer>
<script type="text/javascript">
var varAutoFirePV = 1;
var varClickTracking = 0;
var varCustomerTracking = 1;
var Route = "route_id";
var Ctrl = "control_id";
</script>
<script type="speculationrules">
{"prefetch":[{"source":"document","where":{"and":[{"href_matches":"/en-us/research/*"},{"not":{"href_matches":["/en-us/research/wp-*.php","/en-us/research/wp-admin/*","/en-us/research/wp-content/uploads/*","/en-us/research/wp-content/*","/en-us/research/wp-content/plugins/*","/en-us/research/wp-content/themes/microsoft-research-theme/*","/en-us/research/*\\?(.+)"]}},{"not":{"selector_matches":"a[rel~=\"nofollow\"]"}},{"not":{"selector_matches":".no-prefetch, .no-prefetch a"}}]},"eagerness":"conservative"}]}
</script>
<script>
( function () {
'use strict';
var EMBED_ORIGIN = "https:\/\/www.youtube-nocookie.com";
var CONSENT_MSG = "Video playback requires cookie consent";
// One entry per activated player: { el, id, title, videoId, completed }.
var players = [];
var boundListen = false;
// Cached oEmbed markup may predate the hidden overlay attribute. Keep
// the consent message out of view until WCP resolves authoritatively.
document.querySelectorAll( '.yt-consent-placeholder' ).forEach( function ( placeholder ) {
var overlay = placeholder.querySelector( '.yt-consent-placeholder__overlay' );
if ( overlay && ! placeholder.classList.contains( 'is-consent-required' ) ) {
overlay.hidden = true;
}
} );
/* -----------------------------------------------------------------
* Consent
* ----------------------------------------------------------------- */
function promptForConsent() {
var api = window.siteConsent;
if ( api && typeof api.manageConsent === 'function' ) {
api.manageConsent();
}
}
/* -----------------------------------------------------------------
* 1DS reporting
* ----------------------------------------------------------------- */
function pageName() {
if ( typeof analytics !== 'undefined' && analytics.config ) {
var waConfig = analytics.config.webAnalyticsConfiguration || {};
var coreData = waConfig.coreData || {};
if ( coreData.pageName ) {
return coreData.pageName;
}
}
return document.title;
}
function report( entry, behavior ) {
var data = {
behavior: behavior,
actionType: 'CL',
contentTags: {
vidnm: entry.title || pageName(),
vidid: entry.videoId || ''
}
};
if ( typeof analytics !== 'undefined' && analytics.capturePageAction ) {
analytics.capturePageAction( entry.el, data );
}
if ( document.body.classList.contains( 'logged-in' ) ) {
console.log( data );
}
}
/* -----------------------------------------------------------------
* postMessage bridge
* ----------------------------------------------------------------- */
/**
* Ask a player to start emitting state events.
*
* Retried a few times because the handshake is dropped if it arrives
* before the player inside the iframe has finished booting.
*/
function subscribe( entry, attempt ) {
attempt = attempt || 0;
if ( ! entry.el.contentWindow || entry.acknowledged || attempt > 4 ) {
return;
}
entry.el.contentWindow.postMessage(
JSON.stringify( { event: 'listening', id: entry.id, channel: 'widget' } ),
EMBED_ORIGIN
);
if ( ! entry.acknowledged ) {
setTimeout( function () {
subscribe( entry, attempt + 1 );
}, 750 );
}
}
function playerFor( source ) {
for ( var i = 0; i < players.length; i++ ) {
if ( players[ i ].el.contentWindow === source ) {
return players[ i ];
}
}
return null;
}
function onMessage( event ) {
if ( event.origin !== EMBED_ORIGIN ) {
return;
}
var entry = playerFor( event.source );
if ( ! entry ) {
return;
}
var payload = event.data;
if ( typeof payload === 'string' ) {
try {
payload = JSON.parse( payload );
} catch ( e ) {
return;
}
}
if ( ! payload || ! payload.info ) {
return;
}
entry.acknowledged = true;
// Title and ID arrive alongside state, so 1DS reports the real
// video rather than the page title.
if ( payload.info.videoData ) {
entry.title = payload.info.videoData.title || entry.title;
entry.videoId = payload.info.videoData.video_id || entry.videoId;
if ( entry.title ) {
entry.el.setAttribute( 'data-bi-cN', entry.title );
entry.el.setAttribute( 'data-bi-id', entry.videoId || '' );
}
}
if ( typeof payload.info.playerState === 'undefined' ) {
return;
}
switch ( payload.info.playerState ) {
case 1:
report( entry, 240 );
break;
case 2:
report( entry, 241 );
break;
case 0:
// YouTube repeats the ended state on replay; report once.
if ( ! entry.completed ) {
entry.completed = true;
report( entry, 245 );
}
break;
}
}
function track( iframe, videoId ) {
// Activation can run more than once during a grant → withdraw →
// grant cycle, so do not stack listeners on the same iframe.
// Without this guard each pass would push another entry sharing
// the same contentWindow.
if ( iframe.__msrYtEntry ) {
subscribe( iframe.__msrYtEntry );
return;
}
var entry = {
el: iframe,
id: 'msr-yt-' + players.length,
title: '',
videoId: videoId || '',
completed: false,
acknowledged: false
};
iframe.__msrYtEntry = entry;
players.push( entry );
if ( ! boundListen ) {
window.addEventListener( 'message', onMessage );
boundListen = true;
}
// Retained on the element so revocation can detach it; an
// anonymous handler would accumulate one listener per cycle.
entry.onLoad = function () {
subscribe( entry );
};
iframe.addEventListener( 'load', entry.onLoad );
// The iframe may already be loaded when re-activated after a
// consent change, in which case the load event will not fire again.
subscribe( entry );
}
/* -----------------------------------------------------------------
* Activation
* ----------------------------------------------------------------- */
/**
* Guarantee the player will talk to us, and that it talks to the
* cookie-free host.
*/
function playableUrl( url ) {
if ( ! url ) {
return '';
}
var parsed;
try {
parsed = new URL( url, window.location.href );
} catch ( error ) {
return '';
}
if ( 'http:' !== parsed.protocol && 'https:' !== parsed.protocol ) {
return '';
}
var host = parsed.hostname.toLowerCase().replace( /^(www\.|m\.)/, '' );
var path = parsed.pathname || '';
var isEmbed = ( 'youtube.com' === host || 'youtube-nocookie.com' === host ) &&
0 === path.indexOf( '/embed/' );
var isLiveChat = 'youtube.com' === host && 0 === path.indexOf( '/live_chat' );
if ( ! isEmbed && ! isLiveChat ) {
return '';
}
// youtube-nocookie.com serves no /live_chat, which is why the
// server-side gate leaves chat on youtube.com. Rewriting the host
// here would turn a working chat pane into a 404, so mirror that
// exemption rather than rewriting unconditionally.
if ( isLiveChat ) {
return url;
}
url = url.replace( /^(https?:)?\/\/(www\.|m\.)?youtube\.com\//i, EMBED_ORIGIN + '/' );
// enablejsapi only means anything to the player, and appending it
// to a non-player URL just invites a cache miss.
if ( -1 !== url.indexOf( '/embed/' ) && ! /[?&]enablejsapi=/.test( url ) ) {
url += ( url.indexOf( '?' ) === -1 ? '?' : '&' ) + 'enablejsapi=1';
}
return url;
}
/**
* Reveal every gated embed on the page and begin tracking it.
*
* Exposed globally because consent-manager.js dispatches vendors by name
* from the cookie manifest.
*/
window.youtubeTracking = function () {
// Deliberately not guarded by a page-global "already ran" flag:
// content injected after the first pass (REST/AJAX) must still be
// able to activate. Every step below is idempotent instead.
window._ytTrackingInitialized = true;
document.querySelectorAll( '.yt-consent-placeholder iframe' ).forEach( function ( iframe ) {
var placeholder = iframe.closest( '.yt-consent-placeholder' );
var loading = placeholder && placeholder.querySelector( '.yt-consent-placeholder__loading' );
var overlay = placeholder && placeholder.querySelector( '.yt-consent-placeholder__overlay' );
var videoId = iframe.getAttribute( 'data-youtube-video-id' ) ||
( placeholder && placeholder.getAttribute( 'data-video-id' ) ) || '';
var deferredUrl = iframe.getAttribute( 'data-src' );
// wp_kses can strip data-src from legacy cached oEmbeds after
// the server gate wraps them. The wrapper ID remains inert and
// is enough to reconstruct the nocookie player after consent.
if ( ! deferredUrl && videoId ) {
deferredUrl = EMBED_ORIGIN + '/embed/' + encodeURIComponent( videoId ) + '?enablejsapi=1&rel=0';
}
var next = playableUrl( deferredUrl );
if ( ! next ) {
return;
}
if ( overlay ) {
overlay.hidden = true;
}
if ( placeholder ) {
placeholder.removeAttribute( 'role' );
placeholder.removeAttribute( 'aria-label' );
placeholder.classList.remove( 'is-consent-required' );
placeholder.classList.add( 'is-activated' );
}
// Re-assigning an identical src still reloads the frame, which
// would restart playback on a second pass.
if ( iframe.getAttribute( 'src' ) !== next ) {
if ( loading ) {
loading.hidden = false;
iframe.addEventListener( 'load', function hideLoadingIndicator() {
loading.hidden = true;
iframe.removeEventListener( 'load', hideLoadingIndicator );
} );
}
iframe.src = next;
} else if ( loading ) {
loading.hidden = true;
}
iframe.removeAttribute( 'aria-hidden' );
iframe.removeAttribute( 'tabindex' );
// Live chat is not a player: it never answers the listening
// handshake, and subscribe() would post to the nocookie origin
// while the frame is on youtube.com.
if ( -1 !== next.indexOf( '/embed/' ) ) {
track( iframe, videoId );
}
} );
// Lets the single-video fallback re-run its reachability check now
// that the generic activation loop has assigned the iframe source.
document.dispatchEvent( new CustomEvent( 'msr:youtube-activated' ) );
};
/**
* Tear every embed back down when consent is withdrawn.
*
* Note this cannot remove the .youtube.com cookies themselves — they are
* third-party and HttpOnly, so document.cookie can neither read nor
* expire them. Not dropping them in the first place is the only control
* that works, which is why activation is gated rather than cleaned up.
*/
window.youtubeRevoke = function () {
window._ytTrackingInitialized = false;
players.forEach( function ( entry ) {
try {
entry.el.contentWindow.postMessage(
JSON.stringify( { event: 'command', func: 'stopVideo', args: [] } ),
EMBED_ORIGIN
);
} catch ( e ) {}
entry.el.removeAttribute( 'src' );
entry.el.setAttribute( 'aria-hidden', 'true' );
entry.el.setAttribute( 'tabindex', '-1' );
// Detach the load handler and the element's entry pointer so a
// later re-grant rebinds cleanly instead of stacking listeners.
if ( entry.onLoad ) {
entry.el.removeEventListener( 'load', entry.onLoad );
}
delete entry.el.__msrYtEntry;
} );
players = [];
document.querySelectorAll( '.yt-consent-placeholder iframe' ).forEach( function ( iframe ) {
iframe.removeAttribute( 'src' );
iframe.setAttribute( 'aria-hidden', 'true' );
iframe.setAttribute( 'tabindex', '-1' );
} );
document.querySelectorAll( '.yt-consent-placeholder' ).forEach( function ( el ) {
var loading = el.querySelector( '.yt-consent-placeholder__loading' );
var overlay = el.querySelector( '.yt-consent-placeholder__overlay' );
var consentRequired = window.siteConsent &&
true === window.siteConsent.isConsentRequired;
if ( ! consentRequired ) {
if ( loading ) {
loading.hidden = false;
}
if ( overlay ) {
overlay.hidden = true;
}
el.removeAttribute( 'role' );
el.removeAttribute( 'aria-label' );
el.classList.remove( 'is-activated' );
el.classList.remove( 'is-consent-required' );
return;
}
if ( loading ) {
loading.hidden = true;
}
if ( overlay ) {
overlay.hidden = false;
}
el.setAttribute( 'role', 'region' );
el.setAttribute( 'aria-label', CONSENT_MSG );
el.classList.remove( 'is-activated' );
el.classList.add( 'is-consent-required' );
} );
};
/* -----------------------------------------------------------------
* Wiring
* ----------------------------------------------------------------- */
// Pre-consent, the play button opens the consent dialog rather than
// starting playback. Delegated so it survives overlays being rebuilt.
document.addEventListener( 'click', function ( event ) {
if ( event.target.closest && event.target.closest( '.yt-consent-placeholder__play' ) ) {
promptForConsent();
}
} );
} )();
</script>
<script id="moray_blocks_shared_script-js" src="https://www.microsoft.com/en-us/research/wp-content/plugins/moray-blocks/dist/js/shared.js?ver=0.2.0"></script>
<script id="moray_blocks_frontend_script-js" src="https://www.microsoft.com/en-us/research/wp-content/plugins/moray-blocks/dist/js/frontend.js?ver=0.2.0"></script>
<script id="mwf-moray-js" src="https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/mwf/bundle.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332"></script>
<script id="msr_block_library_plugin_shared-js" src="https://www.microsoft.com/en-us/research/wp-content/plugins/msr-blocks-library/dist/js/shared.js?ver=0.3.0"></script>
<script id="msr_block_library_plugin_frontend-js" src="https://www.microsoft.com/en-us/research/wp-content/plugins/msr-blocks-library/dist/js/frontend.js?ver=ed590bdf264223835ab6"></script>
<script id="msr-accessible-tabs-js" src="https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/accessible-tabs.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332"></script>
<script id="msr-clamp-js" src="https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/clamp.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332"></script>
<script id="msr-responsive-tables-js" src="https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/responsive-tables.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332"></script>
<script id="msr-wedecs-js" src="https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/wedecs.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332"></script>
<script id="wp-dom-ready-js" src="https://www.microsoft.com/en-us/research/wp-includes/js/dist/dom-ready.min.js?ver=a06281ae5cf5500e9317"></script>
<script id="wp-hooks-js" src="https://www.microsoft.com/en-us/research/wp-includes/js/dist/hooks.min.js?ver=7496969728ca0f95732d"></script>
<script id="wp-i18n-js" src="https://www.microsoft.com/en-us/research/wp-includes/js/dist/i18n.min.js?ver=781d11515ad3d91786ec"></script>
<script id="wp-i18n-js-after">
wp.i18n.setLocaleData( { 'text direction\u0004ltr': [ 'ltr' ] } );
//# sourceURL=wp-i18n-js-after
</script>
<script id="wp-a11y-js" src="https://www.microsoft.com/en-us/research/wp-includes/js/dist/a11y.min.js?ver=af934e5259bc51b8718e"></script>
<script id="wp-url-js" src="https://www.microsoft.com/en-us/research/wp-includes/js/dist/url.min.js?ver=bb0f766c3d2efe497871"></script>
<script id="wp-api-fetch-js" src="https://www.microsoft.com/en-us/research/wp-includes/js/dist/api-fetch.min.js?ver=d7efe4dc1468d36c39b8"></script>
<script id="wp-api-fetch-js-after">
wp.apiFetch.use( wp.apiFetch.createRootURLMiddleware( "https://www.microsoft.com/en-us/research/wp-json/" ) );
wp.apiFetch.nonceMiddleware = wp.apiFetch.createNonceMiddleware( "2169b2dac5" );
wp.apiFetch.use( wp.apiFetch.nonceMiddleware );
wp.apiFetch.use( wp.apiFetch.mediaUploadMiddleware );
wp.apiFetch.nonceEndpoint = "https://www.microsoft.com/en-us/research/wp-admin/admin-ajax.php?action=rest-nonce";
//# sourceURL=wp-api-fetch-js-after
</script>
<script id="ms-research-js-extra">
var MSR_i18n = {"currentPageText":"Current Page","pageText":"Page","currentSelections":"Current Selections","currentRemove":"Remove filter for: ","facetSearchUpdate":"Filtering results\u2026","facetSearchDone":"Result filtering completed.","facetErrors":{"dateRange":"Your end date is before your start date. Please check your date range.","dateFormat":"Sorry, we can\u2019t figure out the date range. Try using YYYY-MM-DD formats.","loading":"There was a problem getting the results. You might be offline, or something went wrong in the system. ","autocomplete":"This term has no results and it isn\u2019t in our system."},"expanded":"Expanded","collapsed":"Collapsed","onDemand":{"noResults":"Sorry, there are no items to show with your filters."}};
var MSR_content_refs = {"msr-podcast":"240054"};
var MSRData = {"blogNavigation":{"wrapper":"post-archive-grid","templateId":"post-archive-card","basePaginationUrl":"https://www.microsoft.com/en-us/research/blog/page/","endpointUrl":"https://www.microsoft.com/en-us/research/wp-json/wp/v2/posts/","search":null,"eventTypeTaxonomy":"msr-event-type"}};
var epAutosuggest = {"endpoint":"https://www.microsoft.com/en-us/research/wp-json/microsoft-research/v1/autosuggest","action":"navigate","locale":"en_US"};
//# sourceURL=ms-research-js-extra
</script>
<script id="ms-research-js" src="https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/microsoft-research.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332"></script>
<script id="duet-date-picker-script-js" src="https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/vendor/duet-date-picker/duet/duet.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332"></script>
<script id="microsoft-metrics-consent-js-extra">
var metricsConsentConfig = {"manifest":[{"fn":"facebookTracking","categories":["Analytics","Advertising","SocialMedia"]},{"fn":"linkedinTracking","categories":["Analytics","Advertising","SocialMedia"]},{"fn":"clarityTracking","categories":["Advertising","Analytics"]},{"fn":"youtubeTracking","revoke":"youtubeRevoke","categories":["Analytics","Advertising","SocialMedia"]}],"cookies":{"Advertising":["_clck","_clsk"],"SocialMedia":["bcookie","bscookie","li_gc","lidc","li_sugr","UserMatchHistory","AnalyticsSyncHistory","__Secure-ROLLOUT_TOKEN","__Secure-YNID","VISITOR_PRIVACY_METADATA","__Secure-YEC","VISITOR_INFO1_LIVE","YSC"],"Analytics":["_clck","_clsk","__Secure-ROLLOUT_TOKEN","__Secure-YNID","VISITOR_PRIVACY_METADATA","__Secure-YEC","VISITOR_INFO1_LIVE","YSC"],"AllCategories":["_fbp"]},"mode":"uhf","bannerLocale":"en-us","bannerId":"ms-cookie-banner"};
//# sourceURL=microsoft-metrics-consent-js-extra
</script>
<script id="microsoft-metrics-consent-js" src="https://www.microsoft.com/en-us/research/wp-content/plugins/microsoft-metrics/assets/js/consent-manager.js?ver=1.3.0"></script>
<script id="1DS-init-script-js-after">
<!-- JSLL tracking -->
// 1DS initialization
const analytics = new oneDS.ApplicationInsights();
var gpcOptIn = ( typeof GPC_DataSharingOptIn !== 'undefined' ) ? GPC_DataSharingOptIn : true;
if ( navigator.globalPrivacyControl || document.cookie.includes('3PAdsOptOut=1') ) {
gpcOptIn = false;
}
var config = {
instrumentationKey: "9ec747153cf446f7b4e129dc7eaa8227-f83e8b36-9a56-4437-8103-9010c8e1e72a-6756",
channelConfiguration: {
eventsLimitInMem: 50
},
cookieCfg: {
enabled: true,
domainCookiesEnabled: true,
domain: ".microsoft.com",
},
propertyConfiguration: {
gpcDataSharingOptIn: gpcOptIn,
callback: {
userConsentDetails: ( siteConsent ) ? siteConsent.getConsent : ( typeof WcpConsent !== "undefined" && WcpConsent.siteConsent ) ? WcpConsent.siteConsent.getConsent : undefined
},
},
webAnalyticsConfiguration:{
coreData: {"pageName":"Microsoft Research Forum","pageType":"Event"},
urlCollectQuery: true,
urlCollectHash: true,
autoCapture: {
scroll: true,
pageView: true,
onLoad: true,
onUnload: true,
click: true,
scroll: true,
resize: true,
jsError: true
}
},
customProperties: {
_mkto_trk: function() {
return document.cookie.replace(/(?:(?:^|.*;\s*)_mkto_trk\s*\=\s*([^;]*).*$)|^.*$/, "$1");
}
}
};
// Initialize OneDS SDK
analytics.initialize( config, [] );
//# sourceURL=1DS-init-script-js-after
</script>
<script id="microsoft-uhf-js-extra">
var microsoftUhfSettings = {"homePath":"/en-us/research/","loginUrl":"http://www.microsoft.com/en-us/research/wp-login.php","logoutUrl":"","scripts":[],"inline":["linkedin"]};
//# sourceURL=microsoft-uhf-js-extra
</script>
<script id="microsoft-uhf-js" src="https://www.microsoft.com/en-us/research/wp-content/plugins/microsoft-uhf/assets/microsoft-uhf.js?ver=0.6.1"></script>
<script id="faceted-search-js-extra">
var MSRSearch = {"debugging":"","endpoint":"https://www.microsoft.com/en-us/research/wp-json/microsoft-research/v1/faceted-search","locale":"","origin":"https://www.microsoft.com/en-us/research/","initialLoadResults":{"users":[],"posts":[{"data":{"ID":1171606,"post_author":43868,"post_date":"2026-05-14 10:07:27","post_date_gmt":"2026-05-14 17:07:27","post_content":"\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe true impact of AI does not lie in how well it takes tests or surfs the web, but in how effectively it teaches, coordinates, and operates in a web built for agents rather than humans. We anticipate the augmentation of workflows for AI, reduced communication frictions, and the rise of AI-powered intermediaries that better align markets with human goals. Our research examines how to guide this transition toward open, innovation-driven ecosystems, so that AI\u2019s inevitable advance delivers broad-based welfare gains rather than locking society into narrow, walled gardens.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"level\":3,\"className\":\"h4\"} --\u003E\n\u003Ch3 class=\"wp-block-heading h4\" id=\"explore-more\"\u003EExplore more\u003C/h3\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://cacm.acm.org/opinion/the-agentic-economy/\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EThe Agentic Economy\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.nber.org/system/files/chapters/c15310/c15310.pdf\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EEconomists can help ensure the AI disruption brings society to the good place\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/magentic-marketplace-an-open-source-environment-for-studying-agentic-markets/\"\u003EMagentic Marketplace: An Open-Source Environment for Studying Agentic Markets\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://arxiv.org/abs/2603.25893\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EAgentic Markets: Equilibrium Effects of Improving Consumer Search\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading --\u003E\n\u003Ch2 class=\"wp-block-heading\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EGuiding the AI disruption to the Good Place\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EYASH LARA:\u003C/strong\u003E As agents begin to coordinate tasks, make decisions, and operate across systems, the impact of AI won\u2019t just be technical. It will be economic and social as well.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EJoining us next is David Rothschild, an economist in our New York City lab. David is going to show us how agent-based systems reshape coordination and markets, and how we can guide this transition toward open, innovation-driven ecosystems that benefit society as a whole.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis research looks beyond technical performance to the larger question: how do we ensure AI drives meaningful positive impact at scale?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EDAVID ROTHSCHILD:\u003C/strong\u003E Hi, my name is David Rothschild, and I\u2019m an economist at Microsoft Research in New York City.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EI\u2019m here to talk about a series of papers I\u2019ve been working on with teams in New York and New England in economics and computation. We\u2019re excited about this work because, rather than focusing only on short-term AI impacts, we\u2019re looking more broadly and long term\u2014what happens when AI and agents become ubiquitous? How does that affect markets, society, and all of us?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELet\u2019s start with a puzzle. AI capabilities feel extraordinary, yet large-scale economic disruption still seems muted. Our explanation isn\u2019t about model quality or compute\u2014it\u2019s about how AI is integrated into workflows.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe first key point is that we distinguish three phases of disruption.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn Stage 1, augmentation, AI improves and accelerates individual tasks like writing, summarizing, and coding within workflows designed for humans.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn Stage 2, automation, routine tasks move \u201cunder the hood.\u201d Humans supervise at a high level, but the underlying workflows\u2014forms, approvals, queues\u2014remain largely unchanged. These human-centered structures become bottlenecks that limit further gains.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe real disruption comes in Stage 3: reconstruction. Here, workflows and markets are redesigned from the ground up around AI\u2019s native strengths\u2014parallelism, memory, continuous monitoring, and machine-to-machine interaction. Human-centered interfaces are no longer the constraint.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EMost gains today are still in Stage 1 and early Stage 2. But the largest long-term benefits will come from Stage 3\u2014and getting there is institutionally difficult.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis leads to the second key point: the main barriers to Stage 3 adoption are not technical. Reconstruction requires delegating real authority to agents, building trust and accountability systems, ensuring machine-readable data and constraints, and aligning incentives to reward redesign rather than incremental improvement. It\u2019s an innovator\u2019s dilemma\u2014replacing existing systems with fundamentally new ones.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELike past general-purpose technologies\u2014electricity, computers, the internet\u2014AI has a long adoption curve. The bottleneck lies in organizations and governments, not raw capability.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe third key point is that as we move into Stage 3, coordination becomes central. Communication frictions and system design choices matter more than raw compute.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELower communication barriers and less lock-in enable broader coordination of tasks and payments, increasing returns to innovation and distributing value more equitably. Open ecosystems tend to benefit society, while closed systems concentrate value among incumbents.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis applies to workers as well\u2014their share of value depends on how delegation, monitoring, and access are structured.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe fourth point is more cautionary: the default trajectory favors entrenchment. Incumbents with established user bases are incentivized to maintain control, while building open, safe ecosystems requires costly and complex coordination across organizations and governments.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe result may be acceleration without transformation\u2014AI makes existing platforms faster, but doesn\u2019t fundamentally change them, keeping us stuck in Stage 2.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESo how do we reach a better outcome?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe propose building synthetic agent-based markets to study and influence how transitions to Stage 3 unfold. These environments allow agents representing consumers and businesses to interact under controlled conditions.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThey help us observe dynamics that are difficult to isolate in real-world settings\u2014such as shifts from human-to-agent to agent-to-agent interaction, or what happens when agents are given real authority rather than advisory roles.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe can also study governance mechanisms\u2014constraints, monitoring, auditability\u2014and test resilience by introducing adversarial or malicious actors.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThese experiments make abstract ideas concrete. They show that simply layering AI onto existing workflows yields limited benefits compared to redesigning systems to fully leverage AI capabilities.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAlongside experiments, we develop theoretical models to study these markets at a higher level. These models focus on comparative insights rather than precise predictions\u2014helping us understand how different design choices, such as open vs. closed ecosystems or centralized vs. decentralized coordination, shape outcomes.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe key insight is that architectural choices strongly influence market behavior and value distribution\u2014even when underlying AI capability is held constant.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESo what are the takeaways?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAI disruption is real, but slower than it may appear. Its largest impacts will come from full system reconstruction, not just task-level improvements. Reaching that future requires combining systems thinking, experiments, and theory.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EEarly decisions will matter. The way we design these systems now will shape long-term outcomes for markets, society, and overall welfare.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWaiting is itself a choice\u2014and one with consequences.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThank you.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E","post_title":"Guiding the AI disruption to the Good Place","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"guiding-the-ai-disruption-to-the-good-place","to_ping":"","pinged":"","post_modified":"2026-05-14 10:07:29","post_modified_gmt":"2026-05-14 17:07:29","post_content_filtered":"\n\u003Cp\u003EThe true impact of AI does not lie in how well it takes tests or surfs the web, but in how effectively it teaches, coordinates, and operates in a web built for agents rather than humans. We anticipate the augmentation of workflows for AI, reduced communication frictions, and the rise of AI-powered intermediaries that better align markets with human goals. Our research examines how to guide this transition toward open, innovation-driven ecosystems, so that AI\u2019s inevitable advance delivers broad-based welfare gains rather than locking society into narrow, walled gardens.\u003C/p\u003E\n\n\n\n\u003Ch3 class=\"wp-block-heading h4\" id=\"explore-more\"\u003EExplore more\u003C/h3\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://cacm.acm.org/opinion/the-agentic-economy/\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EThe Agentic Economy\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://www.nber.org/system/files/chapters/c15310/c15310.pdf\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EEconomists can help ensure the AI disruption brings society to the good place\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/magentic-marketplace-an-open-source-environment-for-studying-agentic-markets/\"\u003EMagentic Marketplace: An Open-Source Environment for Studying Agentic Markets\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://arxiv.org/abs/2603.25893\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EAgentic Markets: Equilibrium Effects of Improving Consumer Search\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EGuiding the AI disruption to the Good Place\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EYASH LARA:\u003C/strong\u003E As agents begin to coordinate tasks, make decisions, and operate across systems, the impact of AI won\u2019t just be technical. It will be economic and social as well.\u003C/p\u003E\n\n\n\n\u003Cp\u003EJoining us next is David Rothschild, an economist in our New York City lab. David is going to show us how agent-based systems reshape coordination and markets, and how we can guide this transition toward open, innovation-driven ecosystems that benefit society as a whole.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis research looks beyond technical performance to the larger question: how do we ensure AI drives meaningful positive impact at scale?\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-2\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EDAVID ROTHSCHILD:\u003C/strong\u003E Hi, my name is David Rothschild, and I\u2019m an economist at Microsoft Research in New York City.\u003C/p\u003E\n\n\n\n\u003Cp\u003EI\u2019m here to talk about a series of papers I\u2019ve been working on with teams in New York and New England in economics and computation. We\u2019re excited about this work because, rather than focusing only on short-term AI impacts, we\u2019re looking more broadly and long term\u2014what happens when AI and agents become ubiquitous? How does that affect markets, society, and all of us?\u003C/p\u003E\n\n\n\n\u003Cp\u003ELet\u2019s start with a puzzle. AI capabilities feel extraordinary, yet large-scale economic disruption still seems muted. Our explanation isn\u2019t about model quality or compute\u2014it\u2019s about how AI is integrated into workflows.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe first key point is that we distinguish three phases of disruption.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn Stage 1, augmentation, AI improves and accelerates individual tasks like writing, summarizing, and coding within workflows designed for humans.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn Stage 2, automation, routine tasks move \u201cunder the hood.\u201d Humans supervise at a high level, but the underlying workflows\u2014forms, approvals, queues\u2014remain largely unchanged. These human-centered structures become bottlenecks that limit further gains.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe real disruption comes in Stage 3: reconstruction. Here, workflows and markets are redesigned from the ground up around AI\u2019s native strengths\u2014parallelism, memory, continuous monitoring, and machine-to-machine interaction. Human-centered interfaces are no longer the constraint.\u003C/p\u003E\n\n\n\n\u003Cp\u003EMost gains today are still in Stage 1 and early Stage 2. But the largest long-term benefits will come from Stage 3\u2014and getting there is institutionally difficult.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis leads to the second key point: the main barriers to Stage 3 adoption are not technical. Reconstruction requires delegating real authority to agents, building trust and accountability systems, ensuring machine-readable data and constraints, and aligning incentives to reward redesign rather than incremental improvement. It\u2019s an innovator\u2019s dilemma\u2014replacing existing systems with fundamentally new ones.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELike past general-purpose technologies\u2014electricity, computers, the internet\u2014AI has a long adoption curve. The bottleneck lies in organizations and governments, not raw capability.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe third key point is that as we move into Stage 3, coordination becomes central. Communication frictions and system design choices matter more than raw compute.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELower communication barriers and less lock-in enable broader coordination of tasks and payments, increasing returns to innovation and distributing value more equitably. Open ecosystems tend to benefit society, while closed systems concentrate value among incumbents.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis applies to workers as well\u2014their share of value depends on how delegation, monitoring, and access are structured.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe fourth point is more cautionary: the default trajectory favors entrenchment. Incumbents with established user bases are incentivized to maintain control, while building open, safe ecosystems requires costly and complex coordination across organizations and governments.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe result may be acceleration without transformation\u2014AI makes existing platforms faster, but doesn\u2019t fundamentally change them, keeping us stuck in Stage 2.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo how do we reach a better outcome?\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe propose building synthetic agent-based markets to study and influence how transitions to Stage 3 unfold. These environments allow agents representing consumers and businesses to interact under controlled conditions.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThey help us observe dynamics that are difficult to isolate in real-world settings\u2014such as shifts from human-to-agent to agent-to-agent interaction, or what happens when agents are given real authority rather than advisory roles.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe can also study governance mechanisms\u2014constraints, monitoring, auditability\u2014and test resilience by introducing adversarial or malicious actors.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThese experiments make abstract ideas concrete. They show that simply layering AI onto existing workflows yields limited benefits compared to redesigning systems to fully leverage AI capabilities.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAlongside experiments, we develop theoretical models to study these markets at a higher level. These models focus on comparative insights rather than precise predictions\u2014helping us understand how different design choices, such as open vs. closed ecosystems or centralized vs. decentralized coordination, shape outcomes.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe key insight is that architectural choices strongly influence market behavior and value distribution\u2014even when underlying AI capability is held constant.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo what are the takeaways?\u003C/p\u003E\n\n\n\n\u003Cp\u003EAI disruption is real, but slower than it may appear. Its largest impacts will come from full system reconstruction, not just task-level improvements. Reaching that future requires combining systems thinking, experiments, and theory.\u003C/p\u003E\n\n\n\n\u003Cp\u003EEarly decisions will matter. The way we design these systems now will shape long-term outcomes for markets, society, and overall welfare.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWaiting is itself a choice\u2014and one with consequences.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThank you.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-2\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/guiding-the-ai-disruption-to-the-good-place/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"},{"term_id":13548,"slug":"economics","name":"Economics","parent":0,"term_taxonomy_id":13548,"term_order":0,"facet":"{\"term_id\":13548,\"slug\":\"economics\",\"name\":\"Economics\",\"parent\":0,\"term_taxonomy_id\":13548,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270329,"slug":"season-2-episode-4","name":"Season 2, Episode 4","parent":0,"term_taxonomy_id":270374,"term_order":0,"facet":"{\"term_id\":270329,\"slug\":\"season-2-episode-4\",\"name\":\"Season 2, Episode 4\",\"parent\":0,\"term_taxonomy_id\":270374,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270111,"slug":"ai-for-all","name":"AI for all","parent":0,"term_taxonomy_id":270156,"term_order":0,"facet":"{\"term_id\":270111,\"slug\":\"ai-for-all\",\"name\":\"AI for all\",\"parent\":0,\"term_taxonomy_id\":270156,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"228\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1-480x280.jpg\" class=\"card-img wp-post-image\" alt=\"Microsoft Research Forum S2E4 | David Rothschild | Guiding the AI disruption to the Good Place\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t09:05\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4F_thumbnail_DavidRothschild-1-480x280.jpg\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/guiding-the-ai-disruption-to-the-good-place/\" data-bi-cN=\"Guiding the AI disruption to the Good Place\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled Guiding the AI disruption to the Good Place\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003EGuiding the AI disruption to the Good Place\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tMay 14, 2026\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/yashlara/\"\u003EYash Lara\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/davidmr/\"\u003EDavid Rothschild\u003C/a\u003E\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 4\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"},{"data":{"ID":1171602,"post_author":43868,"post_date":"2026-05-14 10:05:55","post_date_gmt":"2026-05-14 17:05:55","post_content":"\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELanguage models are usually trained to predict the next word, but that does not always lead to the best overall answers. We introduce energy-based fine-tuning, a new method that trains models to produce better full responses, leading to stronger results without the need for complex reward models or verifiers.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"level\":3,\"className\":\"h4\"} --\u003E\n\u003Ch3 class=\"wp-block-heading h4\" id=\"explore-more\"\u003EExplore more\u003C/h3\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/matching-features-not-tokens-energy-based-fine-tuning-of-language-models/\"\u003EMatching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://github.com/sjelassi/ebft_openrlhf\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EAccess code on GitHub\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading --\u003E\n\u003Ch2 class=\"wp-block-heading\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003ENew fine-tuning of language models: Match meaning, not tokens\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EYASH LARA:\u003C/strong\u003E Most language models are still optimized around predicting the next token, even though that doesn\u2019t always lead to the best overall response.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELet\u2019s hear from Carles in our New England lab about energy-based fine-tuning, a different approach that trains models to optimize meaning across an entire response without relying on complex reward models or external verifiers.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIt\u2019s a clean, principled idea with big implications for how we train and deploy models going forward.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOver to you, Carles.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003ECARLES DOMINGO-ENRICH:\u003C/strong\u003E Hi, this is Carles. I\u2019m a Senior Researcher at Microsoft Research New England, and I\u2019ll be talking about energy-based fine-tuning.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis work focuses on training large language models, so I\u2019ll start with an overview of pre-training and post-training.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn pre-training, the most commonly used approach is next-token prediction with cross-entropy loss. In post-training, there are several phases, starting with next-token prediction in the form of mid-training and supervised fine-tuning (SFT), followed by reinforcement learning (RL) fine-tuning\u2014either from human preferences (RLHF) or with verifiable rewards (RLVR).\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELet\u2019s compare next-token prediction with RL using a translation example. The input might be: \u201cTranslate to French: \u2018The cat is sleeping,\u2019\u201d and the output would be \u201cle chat dort.\u201d\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWith next-token prediction, the model is evaluated token by token, and each contributes to the overall loss. With RL, we generate outputs (rollouts), score them with a reward model, and use that signal to update the model.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EBoth approaches have pros and cons. Next-token prediction offers stable training, dense signal, and strong parallelization, but suffers from imitation bias and distribution shift, since it trains only on ground-truth context.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ERL reduces distribution shift by training on model-generated outputs and allows explicit alignment, but it suffers from sparse signal, reduced parallelizability, and requires a reward model.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOur goal is to find a middle ground\u2014an approach that encourages diverse generations, is robust to distribution shifts, provides denser signal than RL, scales well, and does not require a reward model.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOur idea is to use feature maps defined over sequences of tokens. We copy the model we want to train, extract activation values at different layers as features, and define a feature-based moment-matching loss. We then compute rewards from this and optimize using policy gradients.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn this setup, the ground-truth sequence is compared with model-generated outputs using this feature-matching loss.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe feature-matching loss measures how well the model\u2019s distribution matches the ground-truth distribution in an embedding space. We sample context from ground truth and compare the conditional distributions between ground truth and model outputs.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn practice, computing expectations over the full ground-truth distribution is intractable, so we approximate it using available training pairs. Importantly, this approximation preserves the gradients we need.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ENow let\u2019s look at the full algorithm.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EGiven a context like \u201cThe kids were excited because\u2026\u201d and a ground-truth completion such as \u201cit was the last day of school,\u201d we generate multiple candidate completions from the model\u2014for example, \u201cthe summer break was starting,\u201d \u201cthe circus was in town,\u201d or \u201cthe weather was nice.\u201d\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe pass these through a feature network to obtain feature vectors, which are used to compute the feature-matching loss and derive rewards. These rewards are then used to update the model via policy gradients.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELet\u2019s look at results.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EEnergy-based fine-tuning (EBFT) achieves better cross-entropy loss than SFT and RLVR\u2014even though it does not directly optimize for that objective. It also achieves better downstream performance than SFT and is comparable to RLVR, without needing correctness-based rewards.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe feature-matching loss also correlates with cross-entropy but captures long-range calibration across full sequences rather than focusing on individual tokens.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThese results hold across multiple domains, including question answering, coding, and translation. In unstructured coding scenarios, EBFT outperforms both SFT and RLVR in cross-entropy and feature matching, and often matches or exceeds RLVR on downstream tasks.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EI\u2019d like to thank my collaborators and the Microsoft Research environment, which enables high-risk, high-reward research. In this case, that effort has paid off\u2014EBFT is already being used internally at Microsoft to fine-tune models.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe\u2019d love to hear your feedback. Please check out the project repository and website for more details.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThank you for listening.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E","post_title":"New fine-tuning of language models: Match meaning, not tokens","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"new-fine-tuning-of-language-models-match-meaning-not-tokens","to_ping":"","pinged":"","post_modified":"2026-05-14 10:05:57","post_modified_gmt":"2026-05-14 17:05:57","post_content_filtered":"\n\u003Cp\u003ELanguage models are usually trained to predict the next word, but that does not always lead to the best overall answers. We introduce energy-based fine-tuning, a new method that trains models to produce better full responses, leading to stronger results without the need for complex reward models or verifiers.\u003C/p\u003E\n\n\n\n\u003Ch3 class=\"wp-block-heading h4\" id=\"explore-more\"\u003EExplore more\u003C/h3\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/matching-features-not-tokens-energy-based-fine-tuning-of-language-models/\"\u003EMatching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://github.com/sjelassi/ebft_openrlhf\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EAccess code on GitHub\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003ENew fine-tuning of language models: Match meaning, not tokens\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EYASH LARA:\u003C/strong\u003E Most language models are still optimized around predicting the next token, even though that doesn\u2019t always lead to the best overall response.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELet\u2019s hear from Carles in our New England lab about energy-based fine-tuning, a different approach that trains models to optimize meaning across an entire response without relying on complex reward models or external verifiers.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIt\u2019s a clean, principled idea with big implications for how we train and deploy models going forward.\u003C/p\u003E\n\n\n\n\u003Cp\u003EOver to you, Carles.\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-2\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003ECARLES DOMINGO-ENRICH:\u003C/strong\u003E Hi, this is Carles. I\u2019m a Senior Researcher at Microsoft Research New England, and I\u2019ll be talking about energy-based fine-tuning.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis work focuses on training large language models, so I\u2019ll start with an overview of pre-training and post-training.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn pre-training, the most commonly used approach is next-token prediction with cross-entropy loss. In post-training, there are several phases, starting with next-token prediction in the form of mid-training and supervised fine-tuning (SFT), followed by reinforcement learning (RL) fine-tuning\u2014either from human preferences (RLHF) or with verifiable rewards (RLVR).\u003C/p\u003E\n\n\n\n\u003Cp\u003ELet\u2019s compare next-token prediction with RL using a translation example. The input might be: \u201cTranslate to French: \u2018The cat is sleeping,\u2019\u201d and the output would be \u201cle chat dort.\u201d\u003C/p\u003E\n\n\n\n\u003Cp\u003EWith next-token prediction, the model is evaluated token by token, and each contributes to the overall loss. With RL, we generate outputs (rollouts), score them with a reward model, and use that signal to update the model.\u003C/p\u003E\n\n\n\n\u003Cp\u003EBoth approaches have pros and cons. Next-token prediction offers stable training, dense signal, and strong parallelization, but suffers from imitation bias and distribution shift, since it trains only on ground-truth context.\u003C/p\u003E\n\n\n\n\u003Cp\u003ERL reduces distribution shift by training on model-generated outputs and allows explicit alignment, but it suffers from sparse signal, reduced parallelizability, and requires a reward model.\u003C/p\u003E\n\n\n\n\u003Cp\u003EOur goal is to find a middle ground\u2014an approach that encourages diverse generations, is robust to distribution shifts, provides denser signal than RL, scales well, and does not require a reward model.\u003C/p\u003E\n\n\n\n\u003Cp\u003EOur idea is to use feature maps defined over sequences of tokens. We copy the model we want to train, extract activation values at different layers as features, and define a feature-based moment-matching loss. We then compute rewards from this and optimize using policy gradients.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn this setup, the ground-truth sequence is compared with model-generated outputs using this feature-matching loss.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe feature-matching loss measures how well the model\u2019s distribution matches the ground-truth distribution in an embedding space. We sample context from ground truth and compare the conditional distributions between ground truth and model outputs.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn practice, computing expectations over the full ground-truth distribution is intractable, so we approximate it using available training pairs. Importantly, this approximation preserves the gradients we need.\u003C/p\u003E\n\n\n\n\u003Cp\u003ENow let\u2019s look at the full algorithm.\u003C/p\u003E\n\n\n\n\u003Cp\u003EGiven a context like \u201cThe kids were excited because\u2026\u201d and a ground-truth completion such as \u201cit was the last day of school,\u201d we generate multiple candidate completions from the model\u2014for example, \u201cthe summer break was starting,\u201d \u201cthe circus was in town,\u201d or \u201cthe weather was nice.\u201d\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe pass these through a feature network to obtain feature vectors, which are used to compute the feature-matching loss and derive rewards. These rewards are then used to update the model via policy gradients.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELet\u2019s look at results.\u003C/p\u003E\n\n\n\n\u003Cp\u003EEnergy-based fine-tuning (EBFT) achieves better cross-entropy loss than SFT and RLVR\u2014even though it does not directly optimize for that objective. It also achieves better downstream performance than SFT and is comparable to RLVR, without needing correctness-based rewards.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe feature-matching loss also correlates with cross-entropy but captures long-range calibration across full sequences rather than focusing on individual tokens.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThese results hold across multiple domains, including question answering, coding, and translation. In unstructured coding scenarios, EBFT outperforms both SFT and RLVR in cross-entropy and feature matching, and often matches or exceeds RLVR on downstream tasks.\u003C/p\u003E\n\n\n\n\u003Cp\u003EI\u2019d like to thank my collaborators and the Microsoft Research environment, which enables high-risk, high-reward research. In this case, that effort has paid off\u2014EBFT is already being used internally at Microsoft to fine-tune models.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe\u2019d love to hear your feedback. Please check out the project repository and website for more details.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThank you for listening.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-2\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/new-fine-tuning-of-language-models-match-meaning-not-tokens/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270329,"slug":"season-2-episode-4","name":"Season 2, Episode 4","parent":0,"term_taxonomy_id":270374,"term_order":0,"facet":"{\"term_id\":270329,\"slug\":\"season-2-episode-4\",\"name\":\"Season 2, Episode 4\",\"parent\":0,\"term_taxonomy_id\":270374,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270109,"slug":"mission-on-ai","name":"Mission on AI","parent":0,"term_taxonomy_id":270154,"term_order":0,"facet":"{\"term_id\":270109,\"slug\":\"mission-on-ai\",\"name\":\"Mission on AI\",\"parent\":0,\"term_taxonomy_id\":270154,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"228\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1-480x280.jpg\" class=\"card-img wp-post-image\" alt=\"Microsoft Research Forum S2E4 | Carles Domingo-Enrich | New fine-tuning of language models: Match meaning, not tokens\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t07:40\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4E_thumbnail_CarlesDomingoEnrich-1-480x280.jpg\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/new-fine-tuning-of-language-models-match-meaning-not-tokens/\" data-bi-cN=\"New fine-tuning of language models: Match meaning, not tokens\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled New fine-tuning of language models: Match meaning, not tokens\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003ENew fine-tuning of language models: Match meaning, not tokens\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tMay 14, 2026\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/yashlara/\"\u003EYash Lara\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/carlesd/\"\u003ECarles Domingo-Enrich\u003C/a\u003E\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 4\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"},{"data":{"ID":1171596,"post_author":43868,"post_date":"2026-05-14 10:04:14","post_date_gmt":"2026-05-14 17:04:14","post_content":"\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWhat if AI agents could check their work as they go? This verification method extracts verifiable properties from natural language and evaluates them using symbolic or model-based verifiers. Interwhen, a new open-source library, enables real-time verification of each step, helping agents act more safely and reliably in complex, real-world tasks.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"level\":3,\"className\":\"h4\"} --\u003E\n\u003Ch3 class=\"wp-block-heading h4\" id=\"explore-more\"\u003EExplore more\u003C/h3\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/interwhen-a-generalizable-framework-for-verifiable-reasoning-with-test-time-monitors/\" target=\"_blank\" rel=\"noreferrer noopener\"\u003Einterwhen: A Generalizable Framework for Verifiable Reasoning with Test-time Monitors\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"http://github.com/microsoft/interwhen\" target=\"_blank\" rel=\"noreferrer noopener\"\u003Einterwhen on GitHub\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading --\u003E\n\u003Ch2 class=\"wp-block-heading\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EIntroducing Interwhen: Steering reasoning agents with real-time verification\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EYASH LARA:\u003C/strong\u003E As agents take on more responsibility, reliability becomes critical\u2014not just after the fact, but while the agent is working.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EJoining us as a principal researcher from our MSR India Research Lab, Amit Sharma is here to announce Interwhen. This is a new open-source approach to test and verification that lets agents check each step of their own behavior using verifiable properties extracted from natural language.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis work sits right at the intersection of theory and practice, and it tackles one of the hardest problems in generative AI today: knowing when a system is behaving correctly.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOver to you, Amit.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EAMIT SHARMA:\u003C/strong\u003E Hi, I\u2019m Amit Sharma, a Principal Researcher at Microsoft Research in India.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EToday I\u2019m excited to introduce Interwhen, a framework we\u2019ve built for steering reasoning agents with real-time verification. This is joint work with my colleagues at MSR India, as well as Professor Subbarao Kambhampati from Arizona State University.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAI systems today are moving from passive generators of text to agentic systems that can execute actions in the real world, either through tool calls or other means.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFor example, consider an agent that is operating your inbox, executing complex workflows in an organization, or even interacting in the physical world. These agents are equipped with tool calls that can make irreversible changes\u2014for example, sending an email on your behalf, writing to an organizational database, or even moving an actual robot next to you.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESo it is critical not just to verify their final output, but also their intermediate actions.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EA central question for us is: how do we go from agentic execution as it exists today to a future with verified agentic execution?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELet\u2019s look at current verification systems and their limitations. Typically, a verifier has access to the model response at a given time, the task instruction, and possibly some domain policies.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOne approach uses an LLM as a judge, but this is a monolithic verifier and may miss nuanced errors, making it unreliable. Another approach uses formal verifiers, which are reliable but restricted to math and code.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ETo address these challenges, we built Interwhen, a real-time framework for agent verification.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIdeally, we want to use formal verifiers in general domains\u2014but how do we apply them in settings where there may be thousands of tokens and high ambiguity?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe key innovation in Interwhen is an LLM-based projection step that automatically breaks outputs into a list of verifiable properties\u2014desirable conditions that any solution should satisfy.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe framework can then automatically generate a formal specification and create a provable verifier in Python or Lean.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EImportantly, the system does not operate only after the agent is done. It can verify partial responses at any point in time, provide actionable feedback, and steer the reasoning model to avoid violations.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe designed the system with two important principles. First, verification runs asynchronously so the system remains efficient. Second, it is plug-and-play, allowing practitioners to bring in their existing verifiers.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ENow let\u2019s look at how this works in a real-world setting.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EInterwhen achieves state-of-the-art results on agentic benchmarks such as Tau2 Bench, where even small models can rival the accuracy of frontier models.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe first step is extracting verifiable properties\u2014structured, atomic conditions that any solution should satisfy.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFor example, in a retail scenario from Tau2 Bench, a policy expressed in natural language can be converted into properties such as \u201ca refund must go to the original payment method.\u201d From this, the system can automatically generate Python code that acts as a verifier.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis is a one-time operation that can be reviewed and validated. The key value comes at runtime, when variables in the verifier are dynamically filled using the user\u2019s context and the model\u2019s current response.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFor instance, if an agent attempts to process a return using the wrong payment ID, the verifier detects the violation and produces feedback such as \u201cRefund to credit card not allowed.\u201d In the next step, the model uses this feedback to correct its mistake.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EEfficiency is a core design goal. Verification must not only ensure correctness but also be practical.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ETo achieve this, we built a separate monitoring system that tracks the agent\u2019s output in real time and calls verifiers asynchronously for reversible tool calls. Execution is only stopped if an error is detected; otherwise, the process continues as normal.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe framework is general and improves output quality across many tasks, including customer service, travel booking, logical reasoning, and ensuring safety in agent responses.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIt is fully plug-and-play, works with both proprietary and open-weight models, and enforces strict verification\u2014meaning the agent provides no answer if verification fails.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EI\u2019m excited to announce that we have open-sourced Interwhen on GitHub.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAt Microsoft Research, we believe in open collaboration. Our goal is to foster a community focused on AI verification and safety, where we can collectively build and test verifiers for real-world use cases and enable a future where all agentic execution is verified.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIf you\u2019d like to learn more about Interwhen, you can visit the link below.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThank you.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E","post_title":"Introducing Interwhen: Steering reasoning agents with real-time verification","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"introducing-interwhen-steering-reasoning-agents-with-real-time-verification","to_ping":"","pinged":"","post_modified":"2026-05-14 10:11:38","post_modified_gmt":"2026-05-14 17:11:38","post_content_filtered":"\n\u003Cp\u003EWhat if AI agents could check their work as they go? This verification method extracts verifiable properties from natural language and evaluates them using symbolic or model-based verifiers. Interwhen, a new open-source library, enables real-time verification of each step, helping agents act more safely and reliably in complex, real-world tasks.\u003C/p\u003E\n\n\n\n\u003Ch3 class=\"wp-block-heading h4\" id=\"explore-more\"\u003EExplore more\u003C/h3\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/interwhen-a-generalizable-framework-for-verifiable-reasoning-with-test-time-monitors/\" target=\"_blank\" rel=\"noreferrer noopener\"\u003Einterwhen: A Generalizable Framework for Verifiable Reasoning with Test-time Monitors\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"http://github.com/microsoft/interwhen\" target=\"_blank\" rel=\"noopener noreferrer\"\u003Einterwhen on GitHub\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EIntroducing Interwhen: Steering reasoning agents with real-time verification\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EYASH LARA:\u003C/strong\u003E As agents take on more responsibility, reliability becomes critical\u2014not just after the fact, but while the agent is working.\u003C/p\u003E\n\n\n\n\u003Cp\u003EJoining us as a principal researcher from our MSR India Research Lab, Amit Sharma is here to announce Interwhen. This is a new open-source approach to test and verification that lets agents check each step of their own behavior using verifiable properties extracted from natural language.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis work sits right at the intersection of theory and practice, and it tackles one of the hardest problems in generative AI today: knowing when a system is behaving correctly.\u003C/p\u003E\n\n\n\n\u003Cp\u003EOver to you, Amit.\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-2\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EAMIT SHARMA:\u003C/strong\u003E Hi, I\u2019m Amit Sharma, a Principal Researcher at Microsoft Research in India.\u003C/p\u003E\n\n\n\n\u003Cp\u003EToday I\u2019m excited to introduce Interwhen, a framework we\u2019ve built for steering reasoning agents with real-time verification. This is joint work with my colleagues at MSR India, as well as Professor Subbarao Kambhampati from Arizona State University.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAI systems today are moving from passive generators of text to agentic systems that can execute actions in the real world, either through tool calls or other means.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFor example, consider an agent that is operating your inbox, executing complex workflows in an organization, or even interacting in the physical world. These agents are equipped with tool calls that can make irreversible changes\u2014for example, sending an email on your behalf, writing to an organizational database, or even moving an actual robot next to you.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo it is critical not just to verify their final output, but also their intermediate actions.\u003C/p\u003E\n\n\n\n\u003Cp\u003EA central question for us is: how do we go from agentic execution as it exists today to a future with verified agentic execution?\u003C/p\u003E\n\n\n\n\u003Cp\u003ELet\u2019s look at current verification systems and their limitations. Typically, a verifier has access to the model response at a given time, the task instruction, and possibly some domain policies.\u003C/p\u003E\n\n\n\n\u003Cp\u003EOne approach uses an LLM as a judge, but this is a monolithic verifier and may miss nuanced errors, making it unreliable. Another approach uses formal verifiers, which are reliable but restricted to math and code.\u003C/p\u003E\n\n\n\n\u003Cp\u003ETo address these challenges, we built Interwhen, a real-time framework for agent verification.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIdeally, we want to use formal verifiers in general domains\u2014but how do we apply them in settings where there may be thousands of tokens and high ambiguity?\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe key innovation in Interwhen is an LLM-based projection step that automatically breaks outputs into a list of verifiable properties\u2014desirable conditions that any solution should satisfy.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe framework can then automatically generate a formal specification and create a provable verifier in Python or Lean.\u003C/p\u003E\n\n\n\n\u003Cp\u003EImportantly, the system does not operate only after the agent is done. It can verify partial responses at any point in time, provide actionable feedback, and steer the reasoning model to avoid violations.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe designed the system with two important principles. First, verification runs asynchronously so the system remains efficient. Second, it is plug-and-play, allowing practitioners to bring in their existing verifiers.\u003C/p\u003E\n\n\n\n\u003Cp\u003ENow let\u2019s look at how this works in a real-world setting.\u003C/p\u003E\n\n\n\n\u003Cp\u003EInterwhen achieves state-of-the-art results on agentic benchmarks such as Tau2 Bench, where even small models can rival the accuracy of frontier models.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe first step is extracting verifiable properties\u2014structured, atomic conditions that any solution should satisfy.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFor example, in a retail scenario from Tau2 Bench, a policy expressed in natural language can be converted into properties such as \u201ca refund must go to the original payment method.\u201d From this, the system can automatically generate Python code that acts as a verifier.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis is a one-time operation that can be reviewed and validated. The key value comes at runtime, when variables in the verifier are dynamically filled using the user\u2019s context and the model\u2019s current response.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFor instance, if an agent attempts to process a return using the wrong payment ID, the verifier detects the violation and produces feedback such as \u201cRefund to credit card not allowed.\u201d In the next step, the model uses this feedback to correct its mistake.\u003C/p\u003E\n\n\n\n\u003Cp\u003EEfficiency is a core design goal. Verification must not only ensure correctness but also be practical.\u003C/p\u003E\n\n\n\n\u003Cp\u003ETo achieve this, we built a separate monitoring system that tracks the agent\u2019s output in real time and calls verifiers asynchronously for reversible tool calls. Execution is only stopped if an error is detected; otherwise, the process continues as normal.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe framework is general and improves output quality across many tasks, including customer service, travel booking, logical reasoning, and ensuring safety in agent responses.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIt is fully plug-and-play, works with both proprietary and open-weight models, and enforces strict verification\u2014meaning the agent provides no answer if verification fails.\u003C/p\u003E\n\n\n\n\u003Cp\u003EI\u2019m excited to announce that we have open-sourced Interwhen on GitHub.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAt Microsoft Research, we believe in open collaboration. Our goal is to foster a community focused on AI verification and safety, where we can collectively build and test verifiers for real-world use cases and enable a future where all agentic execution is verified.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIf you\u2019d like to learn more about Interwhen, you can visit the link below.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThank you.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-2\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/introducing-interwhen-steering-reasoning-agents-with-real-time-verification/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270329,"slug":"season-2-episode-4","name":"Season 2, Episode 4","parent":0,"term_taxonomy_id":270374,"term_order":0,"facet":"{\"term_id\":270329,\"slug\":\"season-2-episode-4\",\"name\":\"Season 2, Episode 4\",\"parent\":0,\"term_taxonomy_id\":270374,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270112,"slug":"research-exploration","name":"Open research exploration","parent":0,"term_taxonomy_id":270157,"term_order":0,"facet":"{\"term_id\":270112,\"slug\":\"research-exploration\",\"name\":\"Open research exploration\",\"parent\":0,\"term_taxonomy_id\":270157,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"228\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1-480x280.jpg\" class=\"card-img wp-post-image\" alt=\"Microsoft Research Forum S2E4 | Amit Sharma | Introducing Interwhen: Steering reasoning agents with real-time verification\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t06:29\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4D_thumbnail_AmitSharma-1-480x280.jpg\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/introducing-interwhen-steering-reasoning-agents-with-real-time-verification/\" data-bi-cN=\"Introducing Interwhen: Steering reasoning agents with real-time verification\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled Introducing Interwhen: Steering reasoning agents with real-time verification\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003EIntroducing Interwhen: Steering reasoning agents with real-time verification\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tMay 14, 2026\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/yashlara/\"\u003EYash Lara\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/amshar/\"\u003EAmit Sharma\u003C/a\u003E\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 4\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"},{"data":{"ID":1171590,"post_author":43868,"post_date":"2026-05-14 10:02:50","post_date_gmt":"2026-05-14 17:02:50","post_content":"\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWhat if your repo could run itself? GitHub Agentic Workflows bring AI agents directly into repository automation, enabling tasks to run end-to-end inside GitHub Actions. With built-in guardrails and Microsoft-hosted models on Azure, this system introduces a safe, scalable way to automate development workflows using intent-driven AI.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"level\":3,\"className\":\"h4\"} --\u003E\n\u003Ch3 class=\"wp-block-heading h4\" id=\"explore-more\"\u003EExplore more\u003C/h3\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://github.github.com/gh-aw/\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EGitHub Agentic Workflows\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading --\u003E\n\u003Ch2 class=\"wp-block-heading\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EIntroducing GitHub Agentic Workflows: AI that runs your repo\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EYASH LARA:\u003C/strong\u003E What if your repository could run itself, handle issues, automation, and workflows end-to-end without brittle scripts or manual glue?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EPeli, an agentic engineer out of MSR Redmond, is here to announce GitHub agentic workflows. It brings AI agents directly into GitHub Actions, with built-in guardrails and Microsoft-hosted models on Azure.\u003Cbr\u003EThis is a great example of research meeting real developer needs, and doing so in a way that's safe, scalable, and practical.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ETake it away, Peli.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[Music]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EPELI DE HALLEUX:\u003C/strong\u003E Hello everyone. My name is Peli de Halleux and I work for Microsoft Research, and today I'm going to talk about GitHub agentic workflows.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis is a project with GitHub, GitHub Next, and Azure Core, and it aims at using agents to automate all the things. By all the things, we mean automate the entire spectrum of the software development lifecycle, from testing, generating code, documentation, and much more.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe think that automation will be the true accelerator of agentic transformation, going from single developer productivity to 100x, 1000x multiplication. But we also think that this multiplication will come with order and orchestration and processes, and not with swarms and chaos. And this is what we plan to build.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn GitHub agentic workflows, we build agentic human processes powered by GitHub. We have automation through GitHub Actions. We build safety through a sandbox. And we have reasoning through GitHub Copilot CLI. We put the three together, and you have GitHub agentic workflows.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFor example, let\u2019s take a look at this process. It\u2019s a research plan assign. It is a process that involves multiple agents and multiple human intervention.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIt starts with deep research on a schedule that tries to solve a problem, like finding duplicate code. It generates a report. The report is inspected by a developer. The developer decides that the findings are pretty good, and it spawns another agent to turn that report into work items.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ENow, an automated agent, IssueMonster, assigns them to Copilot so that they get turned into code. IssueMonster looks at any issue that's tagged with cookie\u2014it loves to eat cookies.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThen Copilot does the usual transformation to code with a dance with the developer, where the developer can add comments and reviews. Eventually, this becomes a pull request and gets merged.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis is an example of an agentic human process, where multiple agents and multiple humans were involved.\u003Cbr\u003ENow, agents are dangerous if they are left unattended. Whenever you do an agentic process, you may have adversarial strings entering your process through pull requests, issues, or web queries. Anything coming from an untrusted source may try to take over your agents.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe built a sandbox that deals with that. We built a lot of layers of security, and this is one of the aspects that's very important at Microsoft Research. We have the time, the energy, and the people who are able to drill into these problems. This is a very important aspect of agentic workflows: safety.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ETo support the transformation to agents, we also changed the language. We converted the GitHub Action YAML file format to markdown, which is very popular for agents. In our file format, the front matter at the top is the old actions plus some agentic stuff, and then the second part is a prompt.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EBut of course, you don\u2019t edit these files yourself\u2014you use an agent. And you can see on the right that we are using a phone. The future of interacting with agents is going to be natural language-based, and you will not need a laptop. It could be a phone, it could be through voice. And this is also something that we\u2019re exploring in GitHub agentic workflows.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAgentic human processes will not only change software development, they will change any process that uses information and reasoning. And by that, we mean non-developers such as people in marketing, sales, or operations.\u003Cbr\u003EAnybody who\u2019s doing reasoning and using a Copilot today may want to automate this Copilot.\u003Cbr\u003EThis project is a collaboration between Microsoft Research, GitHub Next, GitHub, and Azure Core. The project is open source, and you can find the link below.\u003Cbr\u003EThank you for watching.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E","post_title":"Introducing GitHub Agentic Workflows: AI that runs your repo","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"introducing-github-agentic-workflows-ai-that-runs-your-repo","to_ping":"","pinged":"","post_modified":"2026-05-14 10:04:55","post_modified_gmt":"2026-05-14 17:04:55","post_content_filtered":"\n\u003Cp\u003EWhat if your repo could run itself? GitHub Agentic Workflows bring AI agents directly into repository automation, enabling tasks to run end-to-end inside GitHub Actions. With built-in guardrails and Microsoft-hosted models on Azure, this system introduces a safe, scalable way to automate development workflows using intent-driven AI.\u003C/p\u003E\n\n\n\n\u003Ch3 class=\"wp-block-heading h4\" id=\"explore-more\"\u003EExplore more\u003C/h3\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://github.github.com/gh-aw/\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EGitHub Agentic Workflows\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EIntroducing GitHub Agentic Workflows: AI that runs your repo\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EYASH LARA:\u003C/strong\u003E What if your repository could run itself, handle issues, automation, and workflows end-to-end without brittle scripts or manual glue?\u003C/p\u003E\n\n\n\n\u003Cp\u003EPeli, an agentic engineer out of MSR Redmond, is here to announce GitHub agentic workflows. It brings AI agents directly into GitHub Actions, with built-in guardrails and Microsoft-hosted models on Azure.\u003Cbr\u003EThis is a great example of research meeting real developer needs, and doing so in a way that’s safe, scalable, and practical.\u003C/p\u003E\n\n\n\n\u003Cp\u003ETake it away, Peli.\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-2\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003E[Music]\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EPELI DE HALLEUX:\u003C/strong\u003E Hello everyone. My name is Peli de Halleux and I work for Microsoft Research, and today I’m going to talk about GitHub agentic workflows.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis is a project with GitHub, GitHub Next, and Azure Core, and it aims at using agents to automate all the things. By all the things, we mean automate the entire spectrum of the software development lifecycle, from testing, generating code, documentation, and much more.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe think that automation will be the true accelerator of agentic transformation, going from single developer productivity to 100x, 1000x multiplication. But we also think that this multiplication will come with order and orchestration and processes, and not with swarms and chaos. And this is what we plan to build.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn GitHub agentic workflows, we build agentic human processes powered by GitHub. We have automation through GitHub Actions. We build safety through a sandbox. And we have reasoning through GitHub Copilot CLI. We put the three together, and you have GitHub agentic workflows.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFor example, let\u2019s take a look at this process. It\u2019s a research plan assign. It is a process that involves multiple agents and multiple human intervention.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIt starts with deep research on a schedule that tries to solve a problem, like finding duplicate code. It generates a report. The report is inspected by a developer. The developer decides that the findings are pretty good, and it spawns another agent to turn that report into work items.\u003C/p\u003E\n\n\n\n\u003Cp\u003ENow, an automated agent, IssueMonster, assigns them to Copilot so that they get turned into code. IssueMonster looks at any issue that’s tagged with cookie\u2014it loves to eat cookies.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThen Copilot does the usual transformation to code with a dance with the developer, where the developer can add comments and reviews. Eventually, this becomes a pull request and gets merged.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis is an example of an agentic human process, where multiple agents and multiple humans were involved.\u003Cbr\u003ENow, agents are dangerous if they are left unattended. Whenever you do an agentic process, you may have adversarial strings entering your process through pull requests, issues, or web queries. Anything coming from an untrusted source may try to take over your agents.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe built a sandbox that deals with that. We built a lot of layers of security, and this is one of the aspects that’s very important at Microsoft Research. We have the time, the energy, and the people who are able to drill into these problems. This is a very important aspect of agentic workflows: safety.\u003C/p\u003E\n\n\n\n\u003Cp\u003ETo support the transformation to agents, we also changed the language. We converted the GitHub Action YAML file format to markdown, which is very popular for agents. In our file format, the front matter at the top is the old actions plus some agentic stuff, and then the second part is a prompt.\u003C/p\u003E\n\n\n\n\u003Cp\u003EBut of course, you don\u2019t edit these files yourself\u2014you use an agent. And you can see on the right that we are using a phone. The future of interacting with agents is going to be natural language-based, and you will not need a laptop. It could be a phone, it could be through voice. And this is also something that we\u2019re exploring in GitHub agentic workflows.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAgentic human processes will not only change software development, they will change any process that uses information and reasoning. And by that, we mean non-developers such as people in marketing, sales, or operations.\u003Cbr\u003EAnybody who\u2019s doing reasoning and using a Copilot today may want to automate this Copilot.\u003Cbr\u003EThis project is a collaboration between Microsoft Research, GitHub Next, GitHub, and Azure Core. The project is open source, and you can find the link below.\u003Cbr\u003EThank you for watching.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-2\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/introducing-github-agentic-workflows-ai-that-runs-your-repo/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270329,"slug":"season-2-episode-4","name":"Season 2, Episode 4","parent":0,"term_taxonomy_id":270374,"term_order":0,"facet":"{\"term_id\":270329,\"slug\":\"season-2-episode-4\",\"name\":\"Season 2, Episode 4\",\"parent\":0,\"term_taxonomy_id\":270374,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270110,"slug":"new-advancements-out-of-the-lab","name":"New advancement out of the lab","parent":0,"term_taxonomy_id":270155,"term_order":0,"facet":"{\"term_id\":270110,\"slug\":\"new-advancements-out-of-the-lab\",\"name\":\"New advancement out of the lab\",\"parent\":0,\"term_taxonomy_id\":270155,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"228\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1-480x280.jpg\" class=\"card-img wp-post-image\" alt=\"Microsoft Research Forum S2E4 | Peli de Halleux | Introducing GitHub Agentic Workflows: AI that runs your repo\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t04:43\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4C_thumbnail_PelideHalleux-1-480x280.jpg\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/introducing-github-agentic-workflows-ai-that-runs-your-repo/\" data-bi-cN=\"Introducing GitHub Agentic Workflows: AI that runs your repo\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled Introducing GitHub Agentic Workflows: AI that runs your repo\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003EIntroducing GitHub Agentic Workflows: AI that runs your repo\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tMay 14, 2026\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/yashlara/\"\u003EYash Lara\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/jhalleux/\"\u003EPeli de Halleux\u003C/a\u003E\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 4\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"},{"data":{"ID":1171587,"post_author":43868,"post_date":"2026-05-14 10:00:07","post_date_gmt":"2026-05-14 17:00:07","post_content":"\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWhat if you could run a capable AI agent without leaning on frontier-scale models? MagenticLite is the next generation of Magentic-UI, an agentic experience reimagined and optimized for small language models. It works across both your browser and your local file system in a single workflow, keeping you in the driver's seat at every step. In this session, we'll demo MagenticLite in action and deep dive into the two models powering it: MagenticBrain for planning, coding, and delegation, and Fara-1.5-9B for browser use.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"level\":3,\"className\":\"h4\"} --\u003E\n\u003Ch3 class=\"wp-block-heading h4\" id=\"explore-more\"\u003EExplore more\u003C/h3\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://aka.ms/MagenticBrain-foundry\" type=\"link\" id=\"https://aka.ms/MagenticBrain-foundry\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EMagenticBrain on Microsoft Foundry\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://aka.ms/fara-foundry\" type=\"link\" id=\"https://aka.ms/fara-foundry\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EFara1.5 on Microsoft Foundry\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://aka.ms/MagenticLite\" type=\"link\" id=\"https://aka.ms/MagenticLite\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EMagenticLite on GitHub\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading --\u003E\n\u003Ch2 class=\"wp-block-heading\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EMagenticLite is here: A full-stack agentic experience powered by Small Models\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EYASH LARA:\u003C/strong\u003E Today, I\u2019m excited to announce a new release from the Microsoft Research AI Frontiers Lab, MagenticLite.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EMagenticLite is the next generation of Magentic UI. It works across both the browser and your local file system in a single workflow, keeping you in the driver\u2019s seat at every step.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAlong with MagenticLite, we\u2019re also releasing the two models powering it. MagenticLite is a result of deep collaboration across AI Frontiers\u2014spanning model development, agentic systems, agentic harnesses, and UX. Together, the team has created an end-to-end integration throughout the agentic stack.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe\u2019ll now hear from our amazing colleagues, Weili, Harkirat, and Hussein, who will walk you through MagenticLite and the models, and demo it live.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EWEILI SHI:\u003C/strong\u003E Hi, I\u2019m Weili.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EHUSSEIN MOZANNAR:\u003C/strong\u003E I\u2019m Hussein.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EHARKIRAT BEHL:\u003C/strong\u003E And I\u2019m Harkirat.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe are all members of AI Frontiers, a boutique lab inside Microsoft Research.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EHUSSEIN MOZANNAR:\u003C/strong\u003E And for the past few months, we\u2019ve been working closely together to reimagine what an agentic application can look like\u2014one that\u2019s efficient, capable, and useful for getting work done.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EWEILI SHI:\u003C/strong\u003E To do that, we had to work on the full stack, rethinking everything from how we generate training data, how we design the models, how the harness orchestrates it all, to the user experience that makes the whole thing feel like genuine collaboration.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EHARKIRAT BEHL:\u003C/strong\u003E What really moved the needle was going beyond standard benchmarks. We started with hero use cases\u2014real everyday tasks that people care about. We built our own evals around them and used this signal to drive a flywheel of iterative improvements across the whole stack.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EHUSSEIN MOZANNAR:\u003C/strong\u003E The result today is three joint releases.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFirst is \u003Cstrong\u003EMagenticLite\u003C/strong\u003E, our experience where you can get a real feel of these small-language models to get real work done.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESecond is our computer-use agent model, \u003Cstrong\u003EFara-1.5\u003C/strong\u003E, a state-of-the-art model for its size class.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAnd last but not least, \u003Cstrong\u003EMagentic Orchestrator\u003C/strong\u003E, a model capable of reasoning, delegation, and coding that brings the whole experience together.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ENow, Weili is going to show us how MagenticLite works.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EWEILI SHI:\u003C/strong\u003E Let\u2019s dive into MagenticLite.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EYou may have heard of Magentic UI, the agentic application we released last year that set the foundation for this work. With MagenticLite, we reworked the agent harness to run efficiently on small-language models\u2014making it faster, more lightweight, and no longer reliant on frontier-scale models.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe also refreshed the UX design based on community feedback, making it easier and more natural to work with.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EMagenticLite works across both the browser and your local file system to help you get real work done\u2014whether that\u2019s filling out online forms, making appointments on your behalf, managing files on your desktop, or generating simple code.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELet\u2019s see it in action.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EI can give MagenticLite access to a folder on my computer. Here, I\u2019m giving it notes from the last Microsoft Build conference. I want the agent to search for what has changed and create an update document to help me prepare for the next conference.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIt has successfully accessed my notes and created to-dos. Next, it opened a web browser and started to gather the information I need.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EMagenticLite has access to its own browser running in a virtual machine. This helps minimize the risk of data leakage while allowing Fara, our browser-use model, to operate quickly.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFara performs well at long-running tasks. It looks like Fara has gathered enough information, and the Orchestrator has created a document. Let\u2019s check it out.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EMagenticLite did the job. It included updates on all key sections.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ENext, I\u2019d like to email this document to my colleague. I\u2019ll let MagenticLite do this for me. I can keep working on other things, and MagenticLite will notify me when my attention is needed.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn this case, it needs my help to log into my email account. I\u2019m taking control of the browser to log in.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOnce unblocked, the agent composes the email and can send it once it finishes.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ENow that you\u2019ve seen it in action, let\u2019s take a closer look at the models powering MagenticLite\u2014small models that punch above their weight. Next, Harkirat will introduce Magentic Orchestrator.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EHARKIRAT BEHL:\u003C/strong\u003E Thanks, Weili.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIf MagenticLite is the app you interact with, and Fara drives the browser, then Magentic Orchestrator is the brain that ties it all together. It is the planner, the coder, and the delegator\u2014all in one model.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIts job is to take a messy request\u2014like \u201cBook me a dentist appointment Tuesday afternoon and add it to my calendar\u201d\u2014and convert it into a concrete plan.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe Orchestrator figures out the steps, picks the right tool or sub-agent for each step, writes code when needed, and recovers when something breaks mid-task.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWhat\u2019s interesting is that the recipe is quite simple. Orchestration is usually where people reach for the biggest model they can get, but we wanted to show you that you can push all of this into a small model without giving up capability.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe training is standard SFT, but the key is the data mix\u2014blending complementary styles of data in the right ratio.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe first style is \u003Cstrong\u003Etool-calling data\u003C/strong\u003E\u2014clear requests, selecting tools, calling them with arguments, and handling responses.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe second is \u003Cstrong\u003Eterminal-style data\u003C/strong\u003E, where an agent performs step-by-step actions\u2014observing results and deciding what to do next.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EMixing these teaches the model when to use tools and when to generate code directly.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe result is a model that competes with much larger ones, while staying small enough to run locally alongside Fara\u2014and it\u2019s open weight, so it can integrate into your own systems.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWith that, Hussein will introduce Fara.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EHUSSEIN MOZANNAR:\u003C/strong\u003E One of our goals in AI Frontiers is to train agentic models to complete computer-use tasks. Our bet is that end-to-end synthetic data generation\u2014without human interaction data\u2014can get us there.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELast November, we released Fara-7B. Today, we\u2019re excited to introduce \u003Cstrong\u003EFara-1.5\u003C/strong\u003E, a family of models across three sizes: 4B, 9B, and 27B.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFara-1.5 sets state-of-the-art results for models in its class.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFara can handle real-world web tasks like form filling, booking, shopping, and other repetitive actions. It works by capturing screenshots, analyzing them alongside past context, and predicting the next action\u2014like clicking or typing.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIt operates in a loop: observe, act, evaluate, and continue.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIts action space includes clicking, keyboard input, memory tools, and the ability to ask the user for approval when needed.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOn the Mind2Web benchmark, Fara-1.5 nearly doubles performance compared to Fara-7B, improving from 35% to about 65%.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EBut we didn\u2019t train it just for benchmarks\u2014we trained it for real-world usefulness.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis is enabled by our synthetic data system, \u003Cstrong\u003EFaraGen 2.0\u003C/strong\u003E, which generates training data using:\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003ELive and synthetic web environments\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003EA strong teacher agent\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003EA user simulator\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003EVerification systems for correctness, efficiency, and safety\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis allows us to scale data generation and train models effectively.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELooking ahead, we plan to expand Fara with always-on capabilities, support for additional environments like Windows and Linux, and deeper integration with terminal workflows.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E","post_title":"MagenticLite: A full-stack agentic experience powered by Small Models","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"magenticlite-a-full-stack-agentic-experience-powered-by-small-models","to_ping":"","pinged":"","post_modified":"2026-05-22 05:47:09","post_modified_gmt":"2026-05-22 12:47:09","post_content_filtered":"\n\u003Cp\u003EWhat if you could run a capable AI agent without leaning on frontier-scale models? MagenticLite is the next generation of Magentic-UI, an agentic experience reimagined and optimized for small language models. It works across both your browser and your local file system in a single workflow, keeping you in the driver’s seat at every step. In this session, we’ll demo MagenticLite in action and deep dive into the two models powering it: MagenticBrain for planning, coding, and delegation, and Fara-1.5-9B for browser use.\u003C/p\u003E\n\n\n\n\u003Ch3 class=\"wp-block-heading h4\" id=\"explore-more\"\u003EExplore more\u003C/h3\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://aka.ms/MagenticBrain-foundry\" type=\"link\" id=\"https://aka.ms/MagenticBrain-foundry\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EMagenticBrain on Microsoft Foundry\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://aka.ms/fara-foundry\" type=\"link\" id=\"https://aka.ms/fara-foundry\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EFara1.5 on Microsoft Foundry\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://aka.ms/MagenticLite\" type=\"link\" id=\"https://aka.ms/MagenticLite\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EMagenticLite on GitHub\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EMagenticLite is here: A full-stack agentic experience powered by Small Models\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EYASH LARA:\u003C/strong\u003E Today, I\u2019m excited to announce a new release from the Microsoft Research AI Frontiers Lab, MagenticLite.\u003C/p\u003E\n\n\n\n\u003Cp\u003EMagenticLite is the next generation of Magentic UI. It works across both the browser and your local file system in a single workflow, keeping you in the driver\u2019s seat at every step.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAlong with MagenticLite, we\u2019re also releasing the two models powering it. MagenticLite is a result of deep collaboration across AI Frontiers\u2014spanning model development, agentic systems, agentic harnesses, and UX. Together, the team has created an end-to-end integration throughout the agentic stack.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe\u2019ll now hear from our amazing colleagues, Weili, Harkirat, and Hussein, who will walk you through MagenticLite and the models, and demo it live.\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-2\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003E[MUSIC]\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO\u202fSWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EWEILI SHI:\u003C/strong\u003E Hi, I\u2019m Weili.\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EHUSSEIN MOZANNAR:\u003C/strong\u003E I\u2019m Hussein.\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EHARKIRAT BEHL:\u003C/strong\u003E And I\u2019m Harkirat.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe are all members of AI Frontiers, a boutique lab inside Microsoft Research.\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EHUSSEIN MOZANNAR:\u003C/strong\u003E And for the past few months, we\u2019ve been working closely together to reimagine what an agentic application can look like\u2014one that\u2019s efficient, capable, and useful for getting work done.\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EWEILI SHI:\u003C/strong\u003E To do that, we had to work on the full stack, rethinking everything from how we generate training data, how we design the models, how the harness orchestrates it all, to the user experience that makes the whole thing feel like genuine collaboration.\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EHARKIRAT BEHL:\u003C/strong\u003E What really moved the needle was going beyond standard benchmarks. We started with hero use cases\u2014real everyday tasks that people care about. We built our own evals around them and used this signal to drive a flywheel of iterative improvements across the whole stack.\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EHUSSEIN MOZANNAR:\u003C/strong\u003E The result today is three joint releases.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFirst is \u003Cstrong\u003EMagenticLite\u003C/strong\u003E, our experience where you can get a real feel of these small-language models to get real work done.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESecond is our computer-use agent model, \u003Cstrong\u003EFara-1.5\u003C/strong\u003E, a state-of-the-art model for its size class.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAnd last but not least, \u003Cstrong\u003EMagentic Orchestrator\u003C/strong\u003E, a model capable of reasoning, delegation, and coding that brings the whole experience together.\u003C/p\u003E\n\n\n\n\u003Cp\u003ENow, Weili is going to show us how MagenticLite works.\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EWEILI SHI:\u003C/strong\u003E Let\u2019s dive into MagenticLite.\u003C/p\u003E\n\n\n\n\u003Cp\u003EYou may have heard of Magentic UI, the agentic application we released last year that set the foundation for this work. With MagenticLite, we reworked the agent harness to run efficiently on small-language models\u2014making it faster, more lightweight, and no longer reliant on frontier-scale models.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe also refreshed the UX design based on community feedback, making it easier and more natural to work with.\u003C/p\u003E\n\n\n\n\u003Cp\u003EMagenticLite works across both the browser and your local file system to help you get real work done\u2014whether that\u2019s filling out online forms, making appointments on your behalf, managing files on your desktop, or generating simple code.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELet\u2019s see it in action.\u003C/p\u003E\n\n\n\n\u003Cp\u003EI can give MagenticLite access to a folder on my computer. Here, I\u2019m giving it notes from the last Microsoft Build conference. I want the agent to search for what has changed and create an update document to help me prepare for the next conference.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIt has successfully accessed my notes and created to-dos. Next, it opened a web browser and started to gather the information I need.\u003C/p\u003E\n\n\n\n\u003Cp\u003EMagenticLite has access to its own browser running in a virtual machine. This helps minimize the risk of data leakage while allowing Fara, our browser-use model, to operate quickly.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFara performs well at long-running tasks. It looks like Fara has gathered enough information, and the Orchestrator has created a document. Let\u2019s check it out.\u003C/p\u003E\n\n\n\n\u003Cp\u003EMagenticLite did the job. It included updates on all key sections.\u003C/p\u003E\n\n\n\n\u003Cp\u003ENext, I\u2019d like to email this document to my colleague. I\u2019ll let MagenticLite do this for me. I can keep working on other things, and MagenticLite will notify me when my attention is needed.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn this case, it needs my help to log into my email account. I\u2019m taking control of the browser to log in.\u003C/p\u003E\n\n\n\n\u003Cp\u003EOnce unblocked, the agent composes the email and can send it once it finishes.\u003C/p\u003E\n\n\n\n\u003Cp\u003ENow that you\u2019ve seen it in action, let\u2019s take a closer look at the models powering MagenticLite\u2014small models that punch above their weight. Next, Harkirat will introduce Magentic Orchestrator.\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EHARKIRAT BEHL:\u003C/strong\u003E Thanks, Weili.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIf MagenticLite is the app you interact with, and Fara drives the browser, then Magentic Orchestrator is the brain that ties it all together. It is the planner, the coder, and the delegator\u2014all in one model.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIts job is to take a messy request\u2014like \u201cBook me a dentist appointment Tuesday afternoon and add it to my calendar\u201d\u2014and convert it into a concrete plan.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe Orchestrator figures out the steps, picks the right tool or sub-agent for each step, writes code when needed, and recovers when something breaks mid-task.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWhat\u2019s interesting is that the recipe is quite simple. Orchestration is usually where people reach for the biggest model they can get, but we wanted to show you that you can push all of this into a small model without giving up capability.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe training is standard SFT, but the key is the data mix\u2014blending complementary styles of data in the right ratio.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe first style is \u003Cstrong\u003Etool-calling data\u003C/strong\u003E\u2014clear requests, selecting tools, calling them with arguments, and handling responses.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe second is \u003Cstrong\u003Eterminal-style data\u003C/strong\u003E, where an agent performs step-by-step actions\u2014observing results and deciding what to do next.\u003C/p\u003E\n\n\n\n\u003Cp\u003EMixing these teaches the model when to use tools and when to generate code directly.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe result is a model that competes with much larger ones, while staying small enough to run locally alongside Fara\u2014and it\u2019s open weight, so it can integrate into your own systems.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWith that, Hussein will introduce Fara.\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EHUSSEIN MOZANNAR:\u003C/strong\u003E One of our goals in AI Frontiers is to train agentic models to complete computer-use tasks. Our bet is that end-to-end synthetic data generation\u2014without human interaction data\u2014can get us there.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELast November, we released Fara-7B. Today, we\u2019re excited to introduce \u003Cstrong\u003EFara-1.5\u003C/strong\u003E, a family of models across three sizes: 4B, 9B, and 27B.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFara-1.5 sets state-of-the-art results for models in its class.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFara can handle real-world web tasks like form filling, booking, shopping, and other repetitive actions. It works by capturing screenshots, analyzing them alongside past context, and predicting the next action\u2014like clicking or typing.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIt operates in a loop: observe, act, evaluate, and continue.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIts action space includes clicking, keyboard input, memory tools, and the ability to ask the user for approval when needed.\u003C/p\u003E\n\n\n\n\u003Cp\u003EOn the Mind2Web benchmark, Fara-1.5 nearly doubles performance compared to Fara-7B, improving from 35% to about 65%.\u003C/p\u003E\n\n\n\n\u003Cp\u003EBut we didn\u2019t train it just for benchmarks\u2014we trained it for real-world usefulness.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis is enabled by our synthetic data system, \u003Cstrong\u003EFaraGen 2.0\u003C/strong\u003E, which generates training data using:\u003C/p\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003ELive and synthetic web environments\u003C/li\u003E\n\n\n\n\u003Cli\u003EA strong teacher agent\u003C/li\u003E\n\n\n\n\u003Cli\u003EA user simulator\u003C/li\u003E\n\n\n\n\u003Cli\u003EVerification systems for correctness, efficiency, and safety\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cp\u003EThis allows us to scale data generation and train models effectively.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELooking ahead, we plan to expand Fara with always-on capabilities, support for additional environments like Windows and Linux, and deeper integration with terminal workflows.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-2\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/magenticlite-a-full-stack-agentic-experience-powered-by-small-models/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270329,"slug":"season-2-episode-4","name":"Season 2, Episode 4","parent":0,"term_taxonomy_id":270374,"term_order":0,"facet":"{\"term_id\":270329,\"slug\":\"season-2-episode-4\",\"name\":\"Season 2, Episode 4\",\"parent\":0,\"term_taxonomy_id\":270374,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270110,"slug":"new-advancements-out-of-the-lab","name":"New advancement out of the lab","parent":0,"term_taxonomy_id":270155,"term_order":0,"facet":"{\"term_id\":270110,\"slug\":\"new-advancements-out-of-the-lab\",\"name\":\"New advancement out of the lab\",\"parent\":0,\"term_taxonomy_id\":270155,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"228\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1-480x280.jpg\" class=\"card-img wp-post-image\" alt=\"Microsoft Research Forum S2E4 | Harkirat Behl, Weili Shi, Hussein Mozannar | MagenticLite\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t13:50\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/MRF-S2E4B_thumbnail_Weili_Harkirat_Hussein-1-480x280.jpg\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/magenticlite-a-full-stack-agentic-experience-powered-by-small-models/\" data-bi-cN=\"MagenticLite: A full-stack agentic experience powered by Small Models\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled MagenticLite: A full-stack agentic experience powered by Small Models\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003EMagenticLite: A full-stack agentic experience powered by Small Models\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tMay 14, 2026\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/hbehl/\"\u003EHarkirat Behl\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/weilishi/\"\u003EWeili Shi\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/hmozannar/\"\u003EHussein Mozannar\u003C/a\u003E\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 4\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"},{"data":{"ID":1162548,"post_author":43868,"post_date":"2026-03-03 10:05:21","post_date_gmt":"2026-03-03 18:05:21","post_content":"\u003C!-- wp:group {\"layout\":{\"type\":\"constrained\"}} --\u003E\n\u003Cdiv class=\"wp-block-group\"\u003E\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EDion2 reduces the cost of Muon\u2019s orthonormalization step by orthonormalizing only a small, selected submatrix at each iteration. This lightweight approach preserves Muon\u2019s strong performance while significantly improving scalability of optimizer at scale.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"className\":\"h5\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://github.com/microsoft/dion\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EDion2 on GitHub\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading {\"className\":\"h3\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EDion2: A new simple method to shrink matrix in Muon\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph {\"placeholder\":\"More content...\"} --\u003E\n\u003Cp\u003EWorking out of the New York City lab in the AI frontiers, team Kwangjun is here to introduce Dion2: A simple yet powerful method that makes advanced optimizers more scalable by shrinking the expensive computations they rely on. Dion2 preserves performance while dramatically reducing cost. Opening the door to faster and more flexible training at scale, it's a great example of how elegant focused research can have massive impact on real world AI systems. Let's hear more.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHey, I'm Kwangjun. I'm here to talk about a new simple method to shrink matrix in Muon called Dion2. Our work is in line with AutoML optimizer revolution for context training. AI models consume a lot of compute, and we want to do this better. The current de facto standard for training. A model is something called Adam W and is extremely popular.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAnd here the question is can we do better than this default algorithm? Adam W. Recently there is a new contender for Adam W called Muon which is the orthonormal optimizer. It is the optimizer for matrix parameters which make up nearly all parameters in modern neural networks.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe key idea here is to orthonormalize each update matrix. It has clean theoretical motivation and strong empirical performance. In particular, it has been adopted in frontier models such as Kimi and GLM.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESo should we all switch to Muon? Not so fast, because it turns out it's a bit tricky to scale Muon on to larger scales. It's mostly because Muon relies on matrix level computations, which conflicts with distributed training where weights are sharded and inherently the Muon computation specifically, or normalization, is super linear complexity computation.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis is in contrast with Adam W, which only relies on element wise operations. Hence, it's scalable and compatible with all the distributed training framework. So here's our idea. Can we add a scalability knob to muon to make it more scalable? What I mean by that is let's add a parameter controlling how much of the update matrix we're orthonormalizing.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELet's call it beta, which is in between 0 and 1. When beta is chosen to be 1.0, it recovers the original muon. Hence orthonormalizing using the full matrix, whereas when beta is chosen strictly less than 1.0. We're doing partial orthonormalization, hence leading to cheaper compute and less communication, and hopefully it retains all the benefits of muon.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESo scalability knob again is chosen between 0 and 1. And the goal here is to when the beta scalability factor is chosen strictly less than one. It still preserves strong update quality benefits of muon, even with um with the orthonormalizing only a fraction of the matrix. That's where Dion2 comes in. It's a very simple method to shrink matrix size and muon.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAnd here is the outline of our results. On the left hand side we're showing. Optimizer step benchmark between Muon and Dion2 with various fractions. And you can see that as fraction beta becomes smaller Dion2 achieves speedup over muon. And on the right hand side we're showing the training runs between muon and Dion2. And even with the 25% of orthonormalization, Dion2 is achieving very competitive update quality as muon.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe remarkable thing about Dion2 which makes it practical is its simplicity. It first picks top beta fraction of rows or neurons based on their norms, and then we also orthonormalize only the chosen rows or neurons and decay them the selected rows so that Various diverse neurons get selected across different training steps. This is in contrast with the Dion1, which relied on somewhat complicated low rank approximation for efficiency, and its reliance on linear algebraic function comes with overhead at small scale, which makes it less practical.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAlso, in our experience, the ultra achieves better update quality than beyond one, making it better and more practical optimizer than the other one. Please try out Dion2 in your code base. You can use the same configuration as muon as long as muon is integrated in your code base. In our experience, beta equal to 0.5 or 0.25 leads to a very competitive algorithm with full muon.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis is our experience while integrating our Dion2 and muon to vibe code base. And you can see that Dion with 0.5 or 0 to 0.25 achieves very competitive update quality. In other words, once Muon is in your code base, Dion2 is plug and play. Here are some references you can check out Dion2 implementation from our Microsoft Dion code base, and you can check out more details about Dion2 from our paper.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph {\"placeholder\":\"More content...\"} --\u003E\n\u003Cp\u003EThank you for listening.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E\u003C/div\u003E\n\u003C!-- /wp:group --\u003E","post_title":"Dion2: A new simple method to shrink matrix in Muon","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"dion2-a-new-simple-method-to-shrink-matrix-in-muon","to_ping":"","pinged":"","post_modified":"2026-03-03 10:05:22","post_modified_gmt":"2026-03-03 18:05:22","post_content_filtered":"\n\u003Cdiv class=\"wp-block-group is-layout-constrained wp-block-group-is-layout-constrained\"\u003E\n\u003Cp\u003EDion2 reduces the cost of Muon\u2019s orthonormalization step by orthonormalizing only a small, selected submatrix at each iteration. This lightweight approach preserves Muon\u2019s strong performance while significantly improving scalability of optimizer at scale.\u003C/p\u003E\n\n\n\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://github.com/microsoft/dion\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EDion2 on GitHub\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EDion2: A new simple method to shrink matrix in Muon\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003EWorking out of the New York City lab in the AI frontiers, team Kwangjun is here to introduce Dion2: A simple yet powerful method that makes advanced optimizers more scalable by shrinking the expensive computations they rely on. Dion2 preserves performance while dramatically reducing cost. Opening the door to faster and more flexible training at scale, it’s a great example of how elegant focused research can have massive impact on real world AI systems. Let’s hear more.\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-19\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003EHey, I’m Kwangjun. I’m here to talk about a new simple method to shrink matrix in Muon called Dion2. Our work is in line with AutoML optimizer revolution for context training. AI models consume a lot of compute, and we want to do this better. The current de facto standard for training. A model is something called Adam W and is extremely popular.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAnd here the question is can we do better than this default algorithm? Adam W. Recently there is a new contender for Adam W called Muon which is the orthonormal optimizer. It is the optimizer for matrix parameters which make up nearly all parameters in modern neural networks.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe key idea here is to orthonormalize each update matrix. It has clean theoretical motivation and strong empirical performance. In particular, it has been adopted in frontier models such as Kimi and GLM.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo should we all switch to Muon? Not so fast, because it turns out it’s a bit tricky to scale Muon on to larger scales. It’s mostly because Muon relies on matrix level computations, which conflicts with distributed training where weights are sharded and inherently the Muon computation specifically, or normalization, is super linear complexity computation.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis is in contrast with Adam W, which only relies on element wise operations. Hence, it’s scalable and compatible with all the distributed training framework. So here’s our idea. Can we add a scalability knob to muon to make it more scalable? What I mean by that is let’s add a parameter controlling how much of the update matrix we’re orthonormalizing.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELet’s call it beta, which is in between 0 and 1. When beta is chosen to be 1.0, it recovers the original muon. Hence orthonormalizing using the full matrix, whereas when beta is chosen strictly less than 1.0. We’re doing partial orthonormalization, hence leading to cheaper compute and less communication, and hopefully it retains all the benefits of muon.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo scalability knob again is chosen between 0 and 1. And the goal here is to when the beta scalability factor is chosen strictly less than one. It still preserves strong update quality benefits of muon, even with um with the orthonormalizing only a fraction of the matrix. That’s where Dion2 comes in. It’s a very simple method to shrink matrix size and muon.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAnd here is the outline of our results. On the left hand side we’re showing. Optimizer step benchmark between Muon and Dion2 with various fractions. And you can see that as fraction beta becomes smaller Dion2 achieves speedup over muon. And on the right hand side we’re showing the training runs between muon and Dion2. And even with the 25% of orthonormalization, Dion2 is achieving very competitive update quality as muon.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe remarkable thing about Dion2 which makes it practical is its simplicity. It first picks top beta fraction of rows or neurons based on their norms, and then we also orthonormalize only the chosen rows or neurons and decay them the selected rows so that Various diverse neurons get selected across different training steps. This is in contrast with the Dion1, which relied on somewhat complicated low rank approximation for efficiency, and its reliance on linear algebraic function comes with overhead at small scale, which makes it less practical.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAlso, in our experience, the ultra achieves better update quality than beyond one, making it better and more practical optimizer than the other one. Please try out Dion2 in your code base. You can use the same configuration as muon as long as muon is integrated in your code base. In our experience, beta equal to 0.5 or 0.25 leads to a very competitive algorithm with full muon.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis is our experience while integrating our Dion2 and muon to vibe code base. And you can see that Dion with 0.5 or 0 to 0.25 achieves very competitive update quality. In other words, once Muon is in your code base, Dion2 is plug and play. Here are some references you can check out Dion2 implementation from our Microsoft Dion code base, and you can check out more details about Dion2 from our paper.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThank you for listening.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-19\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/dion2-a-new-simple-method-to-shrink-matrix-in-muon/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-post-option":[{"term_id":269148,"slug":"approved-for-river","name":"Approved for River","parent":0,"term_taxonomy_id":269193,"term_order":0,"facet":"{\"term_id\":269148,\"slug\":\"approved-for-river\",\"name\":\"Approved for River\",\"parent\":0,\"term_taxonomy_id\":269193,\"term_order\":0}"},{"term_id":269142,"slug":"include-in-river","name":"Include in River","parent":0,"term_taxonomy_id":269187,"term_order":0,"facet":"{\"term_id\":269142,\"slug\":\"include-in-river\",\"name\":\"Include in River\",\"parent\":0,\"term_taxonomy_id\":269187,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270235,"slug":"season-2-episode-3","name":"Season 2, Episode 3","parent":0,"term_taxonomy_id":270280,"term_order":0,"facet":"{\"term_id\":270235,\"slug\":\"season-2-episode-3\",\"name\":\"Season 2, Episode 3\",\"parent\":0,\"term_taxonomy_id\":270280,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270110,"slug":"new-advancements-out-of-the-lab","name":"New advancement out of the lab","parent":0,"term_taxonomy_id":270155,"term_order":0,"facet":"{\"term_id\":270110,\"slug\":\"new-advancements-out-of-the-lab\",\"name\":\"New advancement out of the lab\",\"parent\":0,\"term_taxonomy_id\":270155,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"206\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Kwangjun_01-scaled.png\" class=\"card-img wp-post-image\" alt=\"photo of Kwangjun Ahn during the Microsoft Research Forum\" srcset=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Kwangjun_01-scaled.png 2560w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Kwangjun_01-300x158.png 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Kwangjun_01-1024x540.png 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Kwangjun_01-768x405.png 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Kwangjun_01-1536x810.png 1536w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Kwangjun_01-2048x1080.png 2048w\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t06:42\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Kwangjun_01-scaled.png\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/dion2-a-new-simple-method-to-shrink-matrix-in-muon/\" data-bi-cN=\"Dion2: A new simple method to shrink matrix in Muon\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled Dion2: A new simple method to shrink matrix in Muon\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003EDion2: A new simple method to shrink matrix in Muon\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tMarch 3, 2026\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/ansonho/\"\u003EAnson Ho\u003C/a\u003E, Kwangjun Ahn\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 3\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"},{"data":{"ID":1162546,"post_author":43868,"post_date":"2026-03-03 10:05:18","post_date_gmt":"2026-03-03 18:05:18","post_content":"\u003C!-- wp:group {\"layout\":{\"type\":\"constrained\"}} --\u003E\n\u003Cdiv class=\"wp-block-group\"\u003E\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe present Adaptively Rotated Optimization (ARO), a matrix optimizer that speeds up LLM training by applying updates in a rotated, geometry-aware coordinate system. Guided by new insights on global structures on LLM loss landscapes, ARO treats rotation as a unifying principle for sample efficiency, and proposed a new update policy that is applicable to all model weight matrices. In large scale controlled experiments, ARO consistently outperforms AdamW and orthogonalization-based method, maintaining its gains as models and training budgets scale.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"className\":\"h5\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/aro-a-new-lens-on-matrix-optimization-for-large-models/\"\u003EARO: A New Lens on Matrix Optimization for Large Models\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading {\"className\":\"h3\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EARO: A new lens on matrix optimization for LLMs\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph {\"placeholder\":\"More content...\"} --\u003E\n\u003Cp\u003ETraining large language models efficiently is one of the biggest challenges in AI right now.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFrom Microsoft Research, Cambridge. Chao and Wenbo introduced adaptively rotated optimization, or ARO, a new approach to optimization that uses geometry-aware updates to significantly improve training efficiency at scale. This is fresh work coming straight out of the lab, with timely results and big implications for both research and production systems.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHanding it over to you, Chao and Wenbo.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHello everyone. My name is Chao Ma and I'm a senior researcher at Microsoft Research Cambridge. We have been working on innovations in AI efficiency and together with my colleague Wenbo will talk about ARO, a new matrix optimization framework for large models. This has been a collaboration with the amazing team at MSR Cambridge.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ETo begin with, training AI is expensive, and The Optimizers crucially determines how effectively we turn compute into intelligence. The industry so far has been dominated by the use of AdamW for a decade now. Matrix based optimizers like Muon are emerging using gradient localization methods to improve sample efficiency and supporting production scale training.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFrom our perspective, we believe organization as one example of this new optimization research landscape, and the goal of our work is to develop new paradigms in optimization that push the efficiency frontier even further. To explain our approach, let me start with the basics. Gradient descent is the simplest optimizer. It computes the gradient g and use it to update the weights w.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAdaptive optimizers like atom also follows the gradient, but they adaptively rescale the gradient before applying the weight update. The adaptive rescaling is implemented by some nonlinear mapping F, and more recent matrix optimizers like orthogonalization, shampoo and soap, etc. are more complicated.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHowever, despite very different motivations and derivations, they are fundamentally similar to Adam. After some simplifications, we can show that they run Adam like optimization, but in a rotated coordinate system. Concretely, this happens by first rotating the gradient by r transpose, then applied nonlinear map f, and then rotated basis and finally rotating back.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn the literature of matrix optimization, f is typically variance of Adam, and R is chosen as the singular vectors of G.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis reinterpretation is not entirely new, but based on the insights from our research, we believe rotation should not just be a Reinterpretation, but instead a core primitive principle for automatic design. This perspective opens the door to many new ideas in optimization, and in particular, we propose the use of more powerful nonlinear mappings beyond atoms rescaling and more significantly, better rotations that are informed by the choice of F and deliver improved sample efficiency.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAnd that gives our framework adaptively rotated optimization or ARO. Before going to the details of ARO's performance. Let me first comment on why we believe rotations are so important. We hypothesize that it is really rooted in symmetries of loss, landscape, current language models, architectures, choices, for example, transformers and reach rotational symmetries in their weights, which when applied correctly, those rotations will leave the model's predictions unchanged.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EARO exploits this property by moving along those symmetric orbits without changing the model's output. To search for updates are more effective at navigating the loss landscape, and this symmetric hypothesis also enables additional features of error. For example, leveraging cross layer couplings for better performance.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn summary, ARO is a general optimization framework that not only explains many existing matrix optimizers, but also enables the discovery of new update rules that work very well in practice. Now I will hand over to Wenbo to work through our key results.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThanks Chao for introducing ARO. Hi, I\u2019m Wenbo. Since we have covered the key features of the ARO, next I will briefly talk about how it performs in practice to ensure fair comparison. We have carefully built an optimizer benchmark protocol that aims to mitigate many potential forms of evaluation bias.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFirst, we collaborated with Microsoft Research Asia and evaluates error on their newly developed M0 model called Sigma. We measured the optimizer performance using the relative speedup in steps over Adam W. This is shown in y axis, and the x axis indicates the data scale where each unit represents the compute optimal data set sizes.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFor example, this spot indicates the 1.36 times speedup as the eight over trend factor. It means that the ARO only uses about 73% of the data or GPU time to reach the same loss as Adam W at that training stage. Therefore, the higher the speed up, the more efficient the optimizer will be. From the plot we can see the result from two clusters. The lower cluster represents also optimization based methods, including MUA. The output cluster shows two error variations. Their results shows a clear performance scan for error, which persists across data skills. Importantly, ARO can optionally be used as a full model optimizer where all matrix parameters in the model are updated under a single rule, while some also organization methods, although we do not show it here, can be as stable in that setting.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOverall, this suggests ARO converges faster and delivers better sample efficiency under our rigorous benchmarking protocol. Next, we stress test ARO across model scales to check whether its efficiency gain holds as we scale up. We evaluate the arrow on models from 0.3 billion up to 8 billion parameters, covering different architectures and overall training regimes.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis plot shows the speedup comparison of ARO versus muon and Adam w. From this we observe that the arrow remains consistent. Advantages. Around 1.3 and 1.1 times over Adam W and Muon, respectively. As model size increases, we do not see the advantage shrinking at larger scales.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAll of the experiments use our efficient distributed error implementation in our setup of 8 billion rounds. The first step runtime is on par with atom W, meaning that the speed up translates directly into work local time savings. Overall, these results suggest that the ARO speedup is scale robust. Based on this trend, it is likely to maintain its advantages over Adam W and muon as we move on to even larger models. To summarize, ARO is a general framework for matrix optimizer. It not only unifies many existing ones. It also provides design opportunities for novel and more efficient optimizers.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAdditionally, ARO can be easily extended and therefore incorporate new developments in the Matrix Optimizer community for further improvements. Our empirical verifications shows consistent advantages over both Adam W and Muon across different model and data skills. Overall, we believe ARO is a competitive optimizer for training large model at scale.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph {\"placeholder\":\"More content...\"} --\u003E\n\u003Cp\u003EWe want to thank all the collaborators in Microsoft research and all entrants for their valuable contributions, discussions, and feedback. If you are interested, please find the paper link below.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E\u003C/div\u003E\n\u003C!-- /wp:group --\u003E","post_title":"ARO: A new lens on matrix optimization for LLMs","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"aro-a-new-lens-on-matrix-optimization-for-llms","to_ping":"","pinged":"","post_modified":"2026-03-03 10:05:20","post_modified_gmt":"2026-03-03 18:05:20","post_content_filtered":"\n\u003Cdiv class=\"wp-block-group is-layout-constrained wp-block-group-is-layout-constrained\"\u003E\n\u003Cp\u003EWe present Adaptively Rotated Optimization (ARO), a matrix optimizer that speeds up LLM training by applying updates in a rotated, geometry-aware coordinate system. Guided by new insights on global structures on LLM loss landscapes, ARO treats rotation as a unifying principle for sample efficiency, and proposed a new update policy that is applicable to all model weight matrices. In large scale controlled experiments, ARO consistently outperforms AdamW and orthogonalization-based method, maintaining its gains as models and training budgets scale.\u003C/p\u003E\n\n\n\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/aro-a-new-lens-on-matrix-optimization-for-large-models/\"\u003EARO: A New Lens on Matrix Optimization for Large Models\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EARO: A new lens on matrix optimization for LLMs\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003ETraining large language models efficiently is one of the biggest challenges in AI right now.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFrom Microsoft Research, Cambridge. Chao and Wenbo introduced adaptively rotated optimization, or ARO, a new approach to optimization that uses geometry-aware updates to significantly improve training efficiency at scale. This is fresh work coming straight out of the lab, with timely results and big implications for both research and production systems.\u003C/p\u003E\n\n\n\n\u003Cp\u003EHanding it over to you, Chao and Wenbo.\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-20\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003EHello everyone. My name is Chao Ma and I’m a senior researcher at Microsoft Research Cambridge. We have been working on innovations in AI efficiency and together with my colleague Wenbo will talk about ARO, a new matrix optimization framework for large models. This has been a collaboration with the amazing team at MSR Cambridge.\u003C/p\u003E\n\n\n\n\u003Cp\u003ETo begin with, training AI is expensive, and The Optimizers crucially determines how effectively we turn compute into intelligence. The industry so far has been dominated by the use of AdamW for a decade now. Matrix based optimizers like Muon are emerging using gradient localization methods to improve sample efficiency and supporting production scale training.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFrom our perspective, we believe organization as one example of this new optimization research landscape, and the goal of our work is to develop new paradigms in optimization that push the efficiency frontier even further. To explain our approach, let me start with the basics. Gradient descent is the simplest optimizer. It computes the gradient g and use it to update the weights w.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAdaptive optimizers like atom also follows the gradient, but they adaptively rescale the gradient before applying the weight update. The adaptive rescaling is implemented by some nonlinear mapping F, and more recent matrix optimizers like orthogonalization, shampoo and soap, etc. are more complicated.\u003C/p\u003E\n\n\n\n\u003Cp\u003EHowever, despite very different motivations and derivations, they are fundamentally similar to Adam. After some simplifications, we can show that they run Adam like optimization, but in a rotated coordinate system. Concretely, this happens by first rotating the gradient by r transpose, then applied nonlinear map f, and then rotated basis and finally rotating back.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn the literature of matrix optimization, f is typically variance of Adam, and R is chosen as the singular vectors of G.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis reinterpretation is not entirely new, but based on the insights from our research, we believe rotation should not just be a Reinterpretation, but instead a core primitive principle for automatic design. This perspective opens the door to many new ideas in optimization, and in particular, we propose the use of more powerful nonlinear mappings beyond atoms rescaling and more significantly, better rotations that are informed by the choice of F and deliver improved sample efficiency.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAnd that gives our framework adaptively rotated optimization or ARO. Before going to the details of ARO’s performance. Let me first comment on why we believe rotations are so important. We hypothesize that it is really rooted in symmetries of loss, landscape, current language models, architectures, choices, for example, transformers and reach rotational symmetries in their weights, which when applied correctly, those rotations will leave the model’s predictions unchanged.\u003C/p\u003E\n\n\n\n\u003Cp\u003EARO exploits this property by moving along those symmetric orbits without changing the model’s output. To search for updates are more effective at navigating the loss landscape, and this symmetric hypothesis also enables additional features of error. For example, leveraging cross layer couplings for better performance.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn summary, ARO is a general optimization framework that not only explains many existing matrix optimizers, but also enables the discovery of new update rules that work very well in practice. Now I will hand over to Wenbo to work through our key results.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThanks Chao for introducing ARO. Hi, I\u2019m Wenbo. Since we have covered the key features of the ARO, next I will briefly talk about how it performs in practice to ensure fair comparison. We have carefully built an optimizer benchmark protocol that aims to mitigate many potential forms of evaluation bias.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFirst, we collaborated with Microsoft Research Asia and evaluates error on their newly developed M0 model called Sigma. We measured the optimizer performance using the relative speedup in steps over Adam W. This is shown in y axis, and the x axis indicates the data scale where each unit represents the compute optimal data set sizes.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFor example, this spot indicates the 1.36 times speedup as the eight over trend factor. It means that the ARO only uses about 73% of the data or GPU time to reach the same loss as Adam W at that training stage. Therefore, the higher the speed up, the more efficient the optimizer will be. From the plot we can see the result from two clusters. The lower cluster represents also optimization based methods, including MUA. The output cluster shows two error variations. Their results shows a clear performance scan for error, which persists across data skills. Importantly, ARO can optionally be used as a full model optimizer where all matrix parameters in the model are updated under a single rule, while some also organization methods, although we do not show it here, can be as stable in that setting.\u003C/p\u003E\n\n\n\n\u003Cp\u003EOverall, this suggests ARO converges faster and delivers better sample efficiency under our rigorous benchmarking protocol. Next, we stress test ARO across model scales to check whether its efficiency gain holds as we scale up. We evaluate the arrow on models from 0.3 billion up to 8 billion parameters, covering different architectures and overall training regimes.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis plot shows the speedup comparison of ARO versus muon and Adam w. From this we observe that the arrow remains consistent. Advantages. Around 1.3 and 1.1 times over Adam W and Muon, respectively. As model size increases, we do not see the advantage shrinking at larger scales.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAll of the experiments use our efficient distributed error implementation in our setup of 8 billion rounds. The first step runtime is on par with atom W, meaning that the speed up translates directly into work local time savings. Overall, these results suggest that the ARO speedup is scale robust. Based on this trend, it is likely to maintain its advantages over Adam W and muon as we move on to even larger models. To summarize, ARO is a general framework for matrix optimizer. It not only unifies many existing ones. It also provides design opportunities for novel and more efficient optimizers.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAdditionally, ARO can be easily extended and therefore incorporate new developments in the Matrix Optimizer community for further improvements. Our empirical verifications shows consistent advantages over both Adam W and Muon across different model and data skills. Overall, we believe ARO is a competitive optimizer for training large model at scale.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe want to thank all the collaborators in Microsoft research and all entrants for their valuable contributions, discussions, and feedback. If you are interested, please find the paper link below.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-20\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/aro-a-new-lens-on-matrix-optimization-for-llms/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-post-option":[{"term_id":269148,"slug":"approved-for-river","name":"Approved for River","parent":0,"term_taxonomy_id":269193,"term_order":0,"facet":"{\"term_id\":269148,\"slug\":\"approved-for-river\",\"name\":\"Approved for River\",\"parent\":0,\"term_taxonomy_id\":269193,\"term_order\":0}"},{"term_id":269142,"slug":"include-in-river","name":"Include in River","parent":0,"term_taxonomy_id":269187,"term_order":0,"facet":"{\"term_id\":269142,\"slug\":\"include-in-river\",\"name\":\"Include in River\",\"parent\":0,\"term_taxonomy_id\":269187,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270235,"slug":"season-2-episode-3","name":"Season 2, Episode 3","parent":0,"term_taxonomy_id":270280,"term_order":0,"facet":"{\"term_id\":270235,\"slug\":\"season-2-episode-3\",\"name\":\"Season 2, Episode 3\",\"parent\":0,\"term_taxonomy_id\":270280,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270110,"slug":"new-advancements-out-of-the-lab","name":"New advancement out of the lab","parent":0,"term_taxonomy_id":270155,"term_order":0,"facet":"{\"term_id\":270110,\"slug\":\"new-advancements-out-of-the-lab\",\"name\":\"New advancement out of the lab\",\"parent\":0,\"term_taxonomy_id\":270155,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"228\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/WenboGong_ChaoMa-1-480x280.png\" class=\"card-img wp-post-image\" alt=\"photo of Wenbo Gong and Chao Ma during the Microsoft Research Forum\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t08:37\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/WenboGong_ChaoMa-1-480x280.png\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/aro-a-new-lens-on-matrix-optimization-for-llms/\" data-bi-cN=\"ARO: A new lens on matrix optimization for LLMs\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled ARO: A new lens on matrix optimization for LLMs\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003EARO: A new lens on matrix optimization for LLMs\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tMarch 3, 2026\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/ansonho/\"\u003EAnson Ho\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/wenbogong/\"\u003EWenbo Gong\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/chaoma/\"\u003EChao Ma\u003C/a\u003E\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 3\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"},{"data":{"ID":1162544,"post_author":43868,"post_date":"2026-03-03 10:05:16","post_date_gmt":"2026-03-03 18:05:16","post_content":"\u003C!-- wp:group {\"layout\":{\"type\":\"constrained\"}} --\u003E\n\u003Cdiv class=\"wp-block-group\"\u003E\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe Tyger framework enables faster, more accessible medical imaging by streaming raw data to the cloud for accelerated reconstruction\u2014reducing patient wait times and discomfort\u2014while empowering researchers to rapidly test and deploy new algorithms.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"className\":\"h5\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://youtu.be/VEXuQo1VOBc?t=1648\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EASHABot discussed in Satya Nadella\u2019s keynote\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://dl.acm.org/doi/full/10.1145/3706598.3713680\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EASHABot: An LLM-Powered Chatbot to Support the Informational Needs of Community Health Workers\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.youtube.com/watch?v=8HeJ69hZ6Gg\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ECataractBot: An LLM-Powered Experts-in-the-Loop Chatbot for Cataract Patients\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading {\"className\":\"h3\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003ELessons from deploying HealthBots with experts-in-the-loop\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph {\"placeholder\":\"More content...\"} --\u003E\n\u003Cp\u003EIn many parts of the world access to healthcare isn't limited by technology, it is limited by access to human expertise.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EMohit Jain from MSR India will share health bots with expert in the loop systems that combine LLM generated responses with fast expert review deployed through familiar tools like WhatsApp. These systems are already helping patients and frontline healthcare workers get timely, trusted guidance at scale.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis is a great example of research advancing AI for all. Using AI to extend human expertise, not replace it. Mohit, over to you.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHey. Hi, everyone. I'm Mohit Jain. I'm a principal researcher at the Microsoft Research Labs in India. Today I'll be talking about the lesson that I learned while deploying health bots with experts in the loop. Let's get started. So patients, typically those who are undergoing critical treatment, let's say a surgery or cancer treatment, require timely, trustworthy and precise medical information.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHowever, in a country like India with only seven doctors for every 10,000 people. It is very hard for a medical professional to spend a lot of time with every patient, which negatively impacts their information sharing with these patients.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFrom a doctor's perspective, questions are fairly similar across patients. For example, if somebody who is undergoing a cataract surgery or who is going through head and neck cancer treatment might ask, when can I start eating biryani after surgery? Or when can I drive my car again? With the advent of large language models, we somehow assume that it will help us make healthcare more accessible.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHowever, as we all know, they do make errors and they do hallucinate. And because of that. Even organizations like OpenAI and anthropic clearly state that in their usage policy that for critical settings like healthcare, finance or legal, we should not completely rely on large language models. So the main research question that we are trying to answer in this work was that can experts, in this case medical professional, help us improve the current AI systems and in return, can AI system help these medical professionals to attend more and more patients?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESo to answer these research questions, we developed multiple chatbots. One is for patients who are undergoing cataract surgery at an eye hospital. So let's see how it actually works in action. So in this case, the chatbot completely runs on WhatsApp so that there is no requirement to download a new software.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOkay. And also it is most accessible in a country like India. The patient can ask a question in any of the languages that they are comfortable in, and they can even send a voice message. So in this case, they could have easily asked the same question by sending a voice message in Hindi. Here, for the sake of the broader audience, let's say the patient asked this question in English when can I wash my hair after the cataract surgery?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe use a knowledge base that has been curated by the doctors of that eye hospital to generate this answer. Which state you can wash your hair three days after the surgery. However, it comes with a question mark because the answer has still not been verified or vetted by a Doctor. We also need a few related questions so that the patient can continue this conversation\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAt the same time when this answer comes to the patient. The answer is also sent to the doctor who is going to operate on that patient. And along with the question answer pair. The doctor also receives this simple question was the bots answer correct and complete. And the doctor has three options. It can.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThey can say yes, there is a correct answer. They can say no. Or they can even send the question to delegate the question to a patient coordinator to answer it. Because we are doing some kind of classification in the backend, whether it is a medical question, then it goes to a doctor. If it is a logistical question, it goes through a patient coordinator.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELet's say in this case, the doctor says that the answer is incomplete. So that triggers a longer workflow. So we inform the patient upfront that the previously provided answer was invalid. And please wait for the corrected answer. Along with that, we asked the doctor please reply with the correction and hear the beauty of our solution.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe doctor is not required to manually edit the answer. They have to just provide a feedback which can be grammatically incorrect, which can be incomplete within completely informal. In this case, the doctor could have just written two followed with a space and W case, and we would have got it that the doctor is saying that they have to wait for two weeks and not two days.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHowever, that answer is not sent to the patient. The patient receives a more formal version of the answer, saying that better to avoid washing your hair for two weeks after the surgery. And it also comes with a green tick mark and with the statement this is a verified answer by your doctor.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ENow we have this really important data from a expert, a medical expert. So we use that data to update the knowledge base so that the next time if another patient asks a similar question, let's say a different patient or this question surgery is done. Can I shampoo? Which is pretty much the same question but in a different way.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESo we said we inform the patient that previously answer was provided to a similar question. And does this answer your question? Yes or no. So this pre verified answer reduces the expert's workload over a period of time.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ENow this system has evolved into a broader framework and we call that. Build your own expert bot which is available as an open-source tool on the Microsoft GitHub page. So anybody can build this expert in the loop chat bot. We use the same platform to build three of these deployment. One as I told you about for patients who are undergoing cataract surgery at the Sankara Eye Hospital.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis has been running at three hospitals of Sankara Eye in Bangalore, Hyderabad and Jaipur, supporting different languages and different socio economic demography of people. More than 3000 patients have used it in the last 9 to 12 months, and more than 10,000 questions have been answered and more than 20 doctors are actively verifying these answers.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EFor community health workers, we have deployed it with the organization called as Khushi Baby. Right now it is actively working in five different districts in Maharashtra and Rajasthan state of India. More than 11,000 ASHA workers, who are the community health workers in India have used it, and more than 50,000 questions have been answered in the community health worker scenario.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe experts are not doctors, they are actually ASHA supervisors who are the auxiliary nurse midwives and hence, instead of a single doctor in this case, a pool of these ASHA supervisors are the ones who actually get every answer. In case of the third deployment is for patients who are undergoing head and neck cancer surgery at the at the cost of a medical hospital, and that system is called as OncoBot.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESo through all of these deployment, we have we have learned multiple lessons. And the most important ones I am putting it here. The first one is that expert verification. Whenever we say that your doctor doctor, let's say Kaushik has verified your answer, it actually adds a lot of trust to those answers.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESo it's very, very important for us to add this expert in the loop. And also we need to make sure that the workload on the doctor is minimal, and hence that all that shorthand, grammatically incorrect correction is what it needs to make the system work. Also, the system is completely asynchronous. The doctor doesn't have to respond in real time. They can respond whenever they are taking a coffee break or something else.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESecond is that compared to other kind of system where we we really enjoy the creativity of a generative AI system. In our case, reliability bleeds that creativity because the constraint static AI relying on the RAC pipeline actually outperform this fully generic system wherein we can't control for the kind of output it will generate, and that also act as the automatic guardrail to the system.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe third point is that admitting uncertainty, in our case, whenever the bot is unaware or is not sure of an answer, it just says, I don't know. Okay. For example, if a patient would have asked like what should I do in case? Can I do pranayama after my surgery? Which is a kind of foundational yoga, but the knowledge base doesn't have that information, so the bot will just respond saying that I do not know, and the answer will be then given by the doctor.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe fourth is that localization is not translation. A typical Azure Language technology translation works really well well for certain languages, but for local Indic languages like Hindi or Tamil or Telugu, it doesn't work that well because it doesn't have the domain culture or the contextual knowledge.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESo instead of that, we start relying on GPT 2 because we can always provide that context, that this is a question asked by a ASHA worker. For example, when Asha worker asks about the side effects of Antara tika, which is a kind of a contraceptive. Uh, the the translation from the Azure Language technology was Antarctica.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOkay? However, when actually we give this information, this is the question asked by Asha worker, we get the appropriate translation. And finally, in the current ecosystem healthcare ecosystem of India, there is a huge power difference between a doctor and a patient, even between a supervisor and a community health worker.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAnd they are always very hesitant to ask very dumb or simple or basic questions to them. However, because there is a chatbot which is sitting in between the two end users and the chatbot is non-judgmental, and that actually makes the patient or the end user like a community health worker, gives them more power to ask very fundamental or basic questions without getting judged.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph {\"placeholder\":\"More content...\"} --\u003E\n\u003Cp\u003EThat's all from me. Thank you.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E\u003C/div\u003E\n\u003C!-- /wp:group --\u003E","post_title":"Lessons from deploying HealthBots with experts-in-the-loop","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"lessons-from-deploying-healthbots-with-experts-in-the-loop","to_ping":"","pinged":"","post_modified":"2026-03-03 10:05:18","post_modified_gmt":"2026-03-03 18:05:18","post_content_filtered":"\n\u003Cdiv class=\"wp-block-group is-layout-constrained wp-block-group-is-layout-constrained\"\u003E\n\u003Cp\u003EThe Tyger framework enables faster, more accessible medical imaging by streaming raw data to the cloud for accelerated reconstruction\u2014reducing patient wait times and discomfort\u2014while empowering researchers to rapidly test and deploy new algorithms.\u003C/p\u003E\n\n\n\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://youtu.be/VEXuQo1VOBc?t=1648\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EASHABot discussed in Satya Nadella\u2019s keynote\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://dl.acm.org/doi/full/10.1145/3706598.3713680\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EASHABot: An LLM-Powered Chatbot to Support the Informational Needs of Community Health Workers\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://www.youtube.com/watch?v=8HeJ69hZ6Gg\" target=\"_blank\" rel=\"noopener noreferrer\"\u003ECataractBot: An LLM-Powered Experts-in-the-Loop Chatbot for Cataract Patients\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003ELessons from deploying HealthBots with experts-in-the-loop\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn many parts of the world access to healthcare isn’t limited by technology, it is limited by access to human expertise.\u003C/p\u003E\n\n\n\n\u003Cp\u003EMohit Jain from MSR India will share health bots with expert in the loop systems that combine LLM generated responses with fast expert review deployed through familiar tools like WhatsApp. These systems are already helping patients and frontline healthcare workers get timely, trusted guidance at scale.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis is a great example of research advancing AI for all. Using AI to extend human expertise, not replace it. Mohit, over to you.\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-21\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003EHey. Hi, everyone. I’m Mohit Jain. I’m a principal researcher at the Microsoft Research Labs in India. Today I’ll be talking about the lesson that I learned while deploying health bots with experts in the loop. Let’s get started. So patients, typically those who are undergoing critical treatment, let’s say a surgery or cancer treatment, require timely, trustworthy and precise medical information.\u003C/p\u003E\n\n\n\n\u003Cp\u003EHowever, in a country like India with only seven doctors for every 10,000 people. It is very hard for a medical professional to spend a lot of time with every patient, which negatively impacts their information sharing with these patients.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFrom a doctor’s perspective, questions are fairly similar across patients. For example, if somebody who is undergoing a cataract surgery or who is going through head and neck cancer treatment might ask, when can I start eating biryani after surgery? Or when can I drive my car again? With the advent of large language models, we somehow assume that it will help us make healthcare more accessible.\u003C/p\u003E\n\n\n\n\u003Cp\u003EHowever, as we all know, they do make errors and they do hallucinate. And because of that. Even organizations like OpenAI and anthropic clearly state that in their usage policy that for critical settings like healthcare, finance or legal, we should not completely rely on large language models. So the main research question that we are trying to answer in this work was that can experts, in this case medical professional, help us improve the current AI systems and in return, can AI system help these medical professionals to attend more and more patients?\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo to answer these research questions, we developed multiple chatbots. One is for patients who are undergoing cataract surgery at an eye hospital. So let’s see how it actually works in action. So in this case, the chatbot completely runs on WhatsApp so that there is no requirement to download a new software.\u003C/p\u003E\n\n\n\n\u003Cp\u003EOkay. And also it is most accessible in a country like India. The patient can ask a question in any of the languages that they are comfortable in, and they can even send a voice message. So in this case, they could have easily asked the same question by sending a voice message in Hindi. Here, for the sake of the broader audience, let’s say the patient asked this question in English when can I wash my hair after the cataract surgery?\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe use a knowledge base that has been curated by the doctors of that eye hospital to generate this answer. Which state you can wash your hair three days after the surgery. However, it comes with a question mark because the answer has still not been verified or vetted by a Doctor. We also need a few related questions so that the patient can continue this conversation\u003C/p\u003E\n\n\n\n\u003Cp\u003EAt the same time when this answer comes to the patient. The answer is also sent to the doctor who is going to operate on that patient. And along with the question answer pair. The doctor also receives this simple question was the bots answer correct and complete. And the doctor has three options. It can.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThey can say yes, there is a correct answer. They can say no. Or they can even send the question to delegate the question to a patient coordinator to answer it. Because we are doing some kind of classification in the backend, whether it is a medical question, then it goes to a doctor. If it is a logistical question, it goes through a patient coordinator.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELet’s say in this case, the doctor says that the answer is incomplete. So that triggers a longer workflow. So we inform the patient upfront that the previously provided answer was invalid. And please wait for the corrected answer. Along with that, we asked the doctor please reply with the correction and hear the beauty of our solution.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe doctor is not required to manually edit the answer. They have to just provide a feedback which can be grammatically incorrect, which can be incomplete within completely informal. In this case, the doctor could have just written two followed with a space and W case, and we would have got it that the doctor is saying that they have to wait for two weeks and not two days.\u003C/p\u003E\n\n\n\n\u003Cp\u003EHowever, that answer is not sent to the patient. The patient receives a more formal version of the answer, saying that better to avoid washing your hair for two weeks after the surgery. And it also comes with a green tick mark and with the statement this is a verified answer by your doctor.\u003C/p\u003E\n\n\n\n\u003Cp\u003ENow we have this really important data from a expert, a medical expert. So we use that data to update the knowledge base so that the next time if another patient asks a similar question, let’s say a different patient or this question surgery is done. Can I shampoo? Which is pretty much the same question but in a different way.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo we said we inform the patient that previously answer was provided to a similar question. And does this answer your question? Yes or no. So this pre verified answer reduces the expert’s workload over a period of time.\u003C/p\u003E\n\n\n\n\u003Cp\u003ENow this system has evolved into a broader framework and we call that. Build your own expert bot which is available as an open-source tool on the Microsoft GitHub page. So anybody can build this expert in the loop chat bot. We use the same platform to build three of these deployment. One as I told you about for patients who are undergoing cataract surgery at the Sankara Eye Hospital.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis has been running at three hospitals of Sankara Eye in Bangalore, Hyderabad and Jaipur, supporting different languages and different socio economic demography of people. More than 3000 patients have used it in the last 9 to 12 months, and more than 10,000 questions have been answered and more than 20 doctors are actively verifying these answers.\u003C/p\u003E\n\n\n\n\u003Cp\u003EFor community health workers, we have deployed it with the organization called as Khushi Baby. Right now it is actively working in five different districts in Maharashtra and Rajasthan state of India. More than 11,000 ASHA workers, who are the community health workers in India have used it, and more than 50,000 questions have been answered in the community health worker scenario.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe experts are not doctors, they are actually ASHA supervisors who are the auxiliary nurse midwives and hence, instead of a single doctor in this case, a pool of these ASHA supervisors are the ones who actually get every answer. In case of the third deployment is for patients who are undergoing head and neck cancer surgery at the at the cost of a medical hospital, and that system is called as OncoBot.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo through all of these deployment, we have we have learned multiple lessons. And the most important ones I am putting it here. The first one is that expert verification. Whenever we say that your doctor doctor, let’s say Kaushik has verified your answer, it actually adds a lot of trust to those answers.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo it’s very, very important for us to add this expert in the loop. And also we need to make sure that the workload on the doctor is minimal, and hence that all that shorthand, grammatically incorrect correction is what it needs to make the system work. Also, the system is completely asynchronous. The doctor doesn’t have to respond in real time. They can respond whenever they are taking a coffee break or something else.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESecond is that compared to other kind of system where we we really enjoy the creativity of a generative AI system. In our case, reliability bleeds that creativity because the constraint static AI relying on the RAC pipeline actually outperform this fully generic system wherein we can’t control for the kind of output it will generate, and that also act as the automatic guardrail to the system.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe third point is that admitting uncertainty, in our case, whenever the bot is unaware or is not sure of an answer, it just says, I don’t know. Okay. For example, if a patient would have asked like what should I do in case? Can I do pranayama after my surgery? Which is a kind of foundational yoga, but the knowledge base doesn’t have that information, so the bot will just respond saying that I do not know, and the answer will be then given by the doctor.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe fourth is that localization is not translation. A typical Azure Language technology translation works really well well for certain languages, but for local Indic languages like Hindi or Tamil or Telugu, it doesn’t work that well because it doesn’t have the domain culture or the contextual knowledge.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo instead of that, we start relying on GPT 2 because we can always provide that context, that this is a question asked by a ASHA worker. For example, when Asha worker asks about the side effects of Antara tika, which is a kind of a contraceptive. Uh, the the translation from the Azure Language technology was Antarctica.\u003C/p\u003E\n\n\n\n\u003Cp\u003EOkay? However, when actually we give this information, this is the question asked by Asha worker, we get the appropriate translation. And finally, in the current ecosystem healthcare ecosystem of India, there is a huge power difference between a doctor and a patient, even between a supervisor and a community health worker.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAnd they are always very hesitant to ask very dumb or simple or basic questions to them. However, because there is a chatbot which is sitting in between the two end users and the chatbot is non-judgmental, and that actually makes the patient or the end user like a community health worker, gives them more power to ask very fundamental or basic questions without getting judged.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThat’s all from me. Thank you.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-21\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/lessons-from-deploying-healthbots-with-experts-in-the-loop/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"},{"term_id":13553,"slug":"medical-health-genomics","name":"Medical, health and genomics","parent":0,"term_taxonomy_id":13553,"term_order":0,"facet":"{\"term_id\":13553,\"slug\":\"medical-health-genomics\",\"name\":\"Medical, health and genomics\",\"parent\":0,\"term_taxonomy_id\":13553,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-post-option":[{"term_id":269148,"slug":"approved-for-river","name":"Approved for River","parent":0,"term_taxonomy_id":269193,"term_order":0,"facet":"{\"term_id\":269148,\"slug\":\"approved-for-river\",\"name\":\"Approved for River\",\"parent\":0,\"term_taxonomy_id\":269193,\"term_order\":0}"},{"term_id":269142,"slug":"include-in-river","name":"Include in River","parent":0,"term_taxonomy_id":269187,"term_order":0,"facet":"{\"term_id\":269142,\"slug\":\"include-in-river\",\"name\":\"Include in River\",\"parent\":0,\"term_taxonomy_id\":269187,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270235,"slug":"season-2-episode-3","name":"Season 2, Episode 3","parent":0,"term_taxonomy_id":270280,"term_order":0,"facet":"{\"term_id\":270235,\"slug\":\"season-2-episode-3\",\"name\":\"Season 2, Episode 3\",\"parent\":0,\"term_taxonomy_id\":270280,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270111,"slug":"ai-for-all","name":"AI for all","parent":0,"term_taxonomy_id":270156,"term_order":0,"facet":"{\"term_id\":270111,\"slug\":\"ai-for-all\",\"name\":\"AI for all\",\"parent\":0,\"term_taxonomy_id\":270156,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"228\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Mohit_01-480x280.png\" class=\"card-img wp-post-image\" alt=\"photo of Mohit Jain during the Microsoft Research Forum\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t09:55\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Mohit_01-480x280.png\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/lessons-from-deploying-healthbots-with-experts-in-the-loop/\" data-bi-cN=\"Lessons from deploying HealthBots with experts-in-the-loop\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled Lessons from deploying HealthBots with experts-in-the-loop\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003ELessons from deploying HealthBots with experts-in-the-loop\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tMarch 3, 2026\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/ansonho/\"\u003EAnson Ho\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/mohja/\"\u003EMohit Jain\u003C/a\u003E\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 3\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"},{"data":{"ID":1162542,"post_author":43868,"post_date":"2026-03-03 10:05:13","post_date_gmt":"2026-03-03 18:05:13","post_content":"\u003C!-- wp:group {\"layout\":{\"type\":\"constrained\"}} --\u003E\n\u003Cdiv class=\"wp-block-group\"\u003E\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOptiMind is a specialized language model that translates natural-language problem descriptions directly into solver-ready mathematical optimization formulations. This removes one of the most expertise-intensive bottlenecks in optimization workflows and makes advanced optimization more accessible.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"className\":\"h5\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/optimind-teaching-llms-to-think-like-optimization-experts/\"\u003EOptiMind: TeachOptiMind: Teaching LLMs to Think Like Optimization Experts\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/blog/optimind-a-small-language-model-with-optimization-expertise/\"\u003EOptiMind: A small language model with optimization expertise\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003EAccess OptiMind: \u003Ca href=\"https://aka.ms/OptiMindCatalog\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EMicrosoft Foundry\u003C/a\u003E | \u003Ca href=\"https://aka.ms/OptiGuideGithub\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EGitHub\u003C/a\u003E | \u003Ca href=\"https://aka.ms/OptiMindHF\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EHugging Face\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading {\"className\":\"h3\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003ETeaching small language models to think like optimization experts with OptiMind\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph {\"placeholder\":\"More content...\"} --\u003E\n\u003Cp\u003EOptimization problems usually start in plain language notes, constraints, and ideas written by humans, turning those into clean, solver ready math is often the hardest part.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EI have the pleasure of working with closely our next speaker, Xinzhi, a research intern at MSR Redmond and a PhD student at the University of Washington. She'll introduce OptiMind, a specialized language model designed to close that gap by translating natural language problem descriptions directly into formal optimization models.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis work pushes the state of the art and AI reasoning, while also improving reliability in systems where correctness really matters. Let's dive into the details and new capabilities of OptiMind.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHi everyone! I'm Xinzhi, a research intern in the Machine Learning and Optimization group here at Microsoft Research, and a PhD student at the University of Washington. Optimization is the engine behind the world's most critical systems, from global supply chains to logistics planning and vehicle routing.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EToday, I want to introduce OptiMind, which is our new 20 billion parameter reasoning model designed to translate natural language problems into a mixed integer linear programing formulations and a corresponding software code. Essentially, it acts like an operations research expert to solve complex tasks in real life.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn this project, which was a collaboration with the amazing ML team here at Microsoft Research, we made two core contributions. First, we develop a way to clean up incredibly noisy optimization data using the expert knowledge. And second, we train a small model that matches the frontier level performance.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe task of OptiMind works as follows A user describes in natural language a problem that involves some decisions that should be optimized. For example, please provide me an optimal production plan to maximize the profits across six months, given the specific capacity constraints and the role of the language model is to generate the mathematical formulation corresponding to the problem and in the form of an executable server code so that the user could run and get an optimal solution.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAnd when we expect the existing datasets and benchmarks in this domain, we found severe data quality issues. This applies to both the training data set and the benchmarks used for testing. And in fact, we saw some training sets with up to 50% of the problematic instances, which include missing constraints, ambiguous descriptions, or wrong solutions.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWhat we could manually clean the test benchmarks to ensure rigorous evaluations. Fixing the massive training sets is much more difficult. Training on the dirty data was like teaching a student with a textbook full of mistakes. So the first research question became, with such noisy data, how can we still train a competitive and reliable model?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ETo solve this, we started by analyzing how models fail when answering the training questions. We noticed that within specific optimization classes, models make the same structural mistakes repeatedly. Take the traveling salesman problem as an example. Models consistently mishandle the software elimination constraints. Specifically, they often incorrectly apply the constraints to the starting node and the result. This generates infeasible and disconnected loops instead of one continuous route.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHowever, we found that if we add specific hints that could explicitly instruct the model how to handle the starting node, it would fix the logic. The model will follow the hint and generate a correct and feasible formulation, and this gives us a key insight. We don't need to manually fix every data point, and instead our experts can identify these failure modes and write a library of expert hints and think of hints as a guard rail.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESimple instructions like enforcing flow conservation or sending the big M number as the maximum variable instead of a fixed number. We warn the model about these specific pitfalls before it even starts generating the solution. So we utilize the hints in two workflows. First, for data cleaning. We first classify the noisy training data and pair it with the specific hints.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe then feed this to a teacher model, which uses the hint to reason correctly and regenerate a clean solution. And finally, apply majority voting to ensure robustness. This use a rigorous and expert quality dataset with minimal human intervention. We then use this clean data to supervise fine tuning our base model, which is GPT OSS 20 billion.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAnd crucially, we repeat this pipeline at inference time. When the user query comes in, we first classify the problem type by identifying it, for example, as a traveling salesman problem. We then retrieve the specific expert hint and feed it with both the user's question and hint into our model to solve the problem.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAnd if the compute budget allows, we go a step further. We utilize the error messages from the solver to perform the multi-tenant self-correction, ensuring the final code is executable and valid. So our results are very compelling on the clean benchmarks IndustryOR, Mamo-Complex and OptMATH, OptiMind consistently outperform the other open-source reasoning models under 32 billion parameters by at least 10%.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAnd when we compare against frontier models like GPT 5 and O4 mini, we see that with just five tons of self-correction, we measure the performance and sometimes even outperform them. And remember, we are achieving this frontier level performance with only 20 billion parameters. The small scale of a model is also critical for local deployments. It allows organizations to keep sensitive supply chain data privacy on their own GPUs, and makes the research reproducible and accessible for the community.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESo, to summarize, we prove that you can train a competitive domain specific model starting from the noisy data by first distilling expert knowledge into reusable hints. And we believe that applying optimization in industrial practice should be a community effort. So we have open source, our model, the cleaned benchmarks and our examples of experiments so the others can build on top of optimized. And looking ahead, we're also working on fully automated pipelines using frontier models so this approach can easily adapt to new areas such as cloud efficiency, calendar scheduling, urban planning, and more. While we focused on standard optimization families, we hope the community will use our methods to explore new and less standard optimization domains.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph {\"placeholder\":\"More content...\"} --\u003E\n\u003Cp\u003ESo I would like to thank all of you for tuning into this talk. Here are some resources if you're interested to learn more about this project and start using the models.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E\u003C/div\u003E\n\u003C!-- /wp:group --\u003E","post_title":"Teaching small language models to think like optimization experts with OptiMind","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"teaching-small-language-models-to-think-like-optimization-experts-with-optimind","to_ping":"","pinged":"","post_modified":"2026-03-03 10:05:15","post_modified_gmt":"2026-03-03 18:05:15","post_content_filtered":"\n\u003Cdiv class=\"wp-block-group is-layout-constrained wp-block-group-is-layout-constrained\"\u003E\n\u003Cp\u003EOptiMind is a specialized language model that translates natural-language problem descriptions directly into solver-ready mathematical optimization formulations. This removes one of the most expertise-intensive bottlenecks in optimization workflows and makes advanced optimization more accessible.\u003C/p\u003E\n\n\n\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/optimind-teaching-llms-to-think-like-optimization-experts/\"\u003EOptiMind: TeachOptiMind: Teaching LLMs to Think Like Optimization Experts\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/blog/optimind-a-small-language-model-with-optimization-expertise/\"\u003EOptiMind: A small language model with optimization expertise\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003EAccess OptiMind: \u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://aka.ms/OptiMindCatalog\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EMicrosoft Foundry\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E | \u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://aka.ms/OptiGuideGithub\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EGitHub\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E | \u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://aka.ms/OptiMindHF\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EHugging Face\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003ETeaching small language models to think like optimization experts with OptiMind\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003EOptimization problems usually start in plain language notes, constraints, and ideas written by humans, turning those into clean, solver ready math is often the hardest part.\u003C/p\u003E\n\n\n\n\u003Cp\u003EI have the pleasure of working with closely our next speaker, Xinzhi, a research intern at MSR Redmond and a PhD student at the University of Washington. She’ll introduce OptiMind, a specialized language model designed to close that gap by translating natural language problem descriptions directly into formal optimization models.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis work pushes the state of the art and AI reasoning, while also improving reliability in systems where correctness really matters. Let’s dive into the details and new capabilities of OptiMind.\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-22\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003EHi everyone! I’m Xinzhi, a research intern in the Machine Learning and Optimization group here at Microsoft Research, and a PhD student at the University of Washington. Optimization is the engine behind the world’s most critical systems, from global supply chains to logistics planning and vehicle routing.\u003C/p\u003E\n\n\n\n\u003Cp\u003EToday, I want to introduce OptiMind, which is our new 20 billion parameter reasoning model designed to translate natural language problems into a mixed integer linear programing formulations and a corresponding software code. Essentially, it acts like an operations research expert to solve complex tasks in real life.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn this project, which was a collaboration with the amazing ML team here at Microsoft Research, we made two core contributions. First, we develop a way to clean up incredibly noisy optimization data using the expert knowledge. And second, we train a small model that matches the frontier level performance.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe task of OptiMind works as follows A user describes in natural language a problem that involves some decisions that should be optimized. For example, please provide me an optimal production plan to maximize the profits across six months, given the specific capacity constraints and the role of the language model is to generate the mathematical formulation corresponding to the problem and in the form of an executable server code so that the user could run and get an optimal solution.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAnd when we expect the existing datasets and benchmarks in this domain, we found severe data quality issues. This applies to both the training data set and the benchmarks used for testing. And in fact, we saw some training sets with up to 50% of the problematic instances, which include missing constraints, ambiguous descriptions, or wrong solutions.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWhat we could manually clean the test benchmarks to ensure rigorous evaluations. Fixing the massive training sets is much more difficult. Training on the dirty data was like teaching a student with a textbook full of mistakes. So the first research question became, with such noisy data, how can we still train a competitive and reliable model?\u003C/p\u003E\n\n\n\n\u003Cp\u003ETo solve this, we started by analyzing how models fail when answering the training questions. We noticed that within specific optimization classes, models make the same structural mistakes repeatedly. Take the traveling salesman problem as an example. Models consistently mishandle the software elimination constraints. Specifically, they often incorrectly apply the constraints to the starting node and the result. This generates infeasible and disconnected loops instead of one continuous route.\u003C/p\u003E\n\n\n\n\u003Cp\u003EHowever, we found that if we add specific hints that could explicitly instruct the model how to handle the starting node, it would fix the logic. The model will follow the hint and generate a correct and feasible formulation, and this gives us a key insight. We don’t need to manually fix every data point, and instead our experts can identify these failure modes and write a library of expert hints and think of hints as a guard rail.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESimple instructions like enforcing flow conservation or sending the big M number as the maximum variable instead of a fixed number. We warn the model about these specific pitfalls before it even starts generating the solution. So we utilize the hints in two workflows. First, for data cleaning. We first classify the noisy training data and pair it with the specific hints.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe then feed this to a teacher model, which uses the hint to reason correctly and regenerate a clean solution. And finally, apply majority voting to ensure robustness. This use a rigorous and expert quality dataset with minimal human intervention. We then use this clean data to supervise fine tuning our base model, which is GPT OSS 20 billion.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAnd crucially, we repeat this pipeline at inference time. When the user query comes in, we first classify the problem type by identifying it, for example, as a traveling salesman problem. We then retrieve the specific expert hint and feed it with both the user’s question and hint into our model to solve the problem.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAnd if the compute budget allows, we go a step further. We utilize the error messages from the solver to perform the multi-tenant self-correction, ensuring the final code is executable and valid. So our results are very compelling on the clean benchmarks IndustryOR, Mamo-Complex and OptMATH, OptiMind consistently outperform the other open-source reasoning models under 32 billion parameters by at least 10%.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAnd when we compare against frontier models like GPT 5 and O4 mini, we see that with just five tons of self-correction, we measure the performance and sometimes even outperform them. And remember, we are achieving this frontier level performance with only 20 billion parameters. The small scale of a model is also critical for local deployments. It allows organizations to keep sensitive supply chain data privacy on their own GPUs, and makes the research reproducible and accessible for the community.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo, to summarize, we prove that you can train a competitive domain specific model starting from the noisy data by first distilling expert knowledge into reusable hints. And we believe that applying optimization in industrial practice should be a community effort. So we have open source, our model, the cleaned benchmarks and our examples of experiments so the others can build on top of optimized. And looking ahead, we’re also working on fully automated pipelines using frontier models so this approach can easily adapt to new areas such as cloud efficiency, calendar scheduling, urban planning, and more. While we focused on standard optimization families, we hope the community will use our methods to explore new and less standard optimization domains.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo I would like to thank all of you for tuning into this talk. Here are some resources if you’re interested to learn more about this project and start using the models.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-22\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/teaching-small-language-models-to-think-like-optimization-experts-with-optimind/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-post-option":[{"term_id":269148,"slug":"approved-for-river","name":"Approved for River","parent":0,"term_taxonomy_id":269193,"term_order":0,"facet":"{\"term_id\":269148,\"slug\":\"approved-for-river\",\"name\":\"Approved for River\",\"parent\":0,\"term_taxonomy_id\":269193,\"term_order\":0}"},{"term_id":269142,"slug":"include-in-river","name":"Include in River","parent":0,"term_taxonomy_id":269187,"term_order":0,"facet":"{\"term_id\":269142,\"slug\":\"include-in-river\",\"name\":\"Include in River\",\"parent\":0,\"term_taxonomy_id\":269187,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270235,"slug":"season-2-episode-3","name":"Season 2, Episode 3","parent":0,"term_taxonomy_id":270280,"term_order":0,"facet":"{\"term_id\":270235,\"slug\":\"season-2-episode-3\",\"name\":\"Season 2, Episode 3\",\"parent\":0,\"term_taxonomy_id\":270280,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270109,"slug":"mission-on-ai","name":"Mission on AI","parent":0,"term_taxonomy_id":270154,"term_order":0,"facet":"{\"term_id\":270109,\"slug\":\"mission-on-ai\",\"name\":\"Mission on AI\",\"parent\":0,\"term_taxonomy_id\":270154,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"228\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Xinzhi_01-480x280.png\" class=\"card-img wp-post-image\" alt=\"photo of Xinzhi Zhang during the Microsoft Research Forum\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t07:55\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Xinzhi_01-480x280.png\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/teaching-small-language-models-to-think-like-optimization-experts-with-optimind/\" data-bi-cN=\"Teaching small language models to think like optimization experts with OptiMind\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled Teaching small language models to think like optimization experts with OptiMind\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003ETeaching small language models to think like optimization experts with OptiMind\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tMarch 3, 2026\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/ansonho/\"\u003EAnson Ho\u003C/a\u003E, \u003Ca href=\"https://www.linkedin.com/in/xinzhi-zhang-55a8662a2/\"\u003EXinzhi Zhang\u003C/a\u003E\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 3\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"},{"data":{"ID":1162539,"post_author":43868,"post_date":"2026-03-03 10:05:11","post_date_gmt":"2026-03-03 18:05:11","post_content":"\u003C!-- wp:group {\"layout\":{\"type\":\"constrained\"}} --\u003E\n\u003Cdiv class=\"wp-block-group\"\u003E\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAgent Lightning is an agent optimization framework that enables agents to learn from their experiences through reinforcement learning and other methods. By treating agents as first-class citizens, optimization becomes automatic for any agent with minimal code changes.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"className\":\"h5\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://github.com/microsoft/agent-lightning\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EAgent Lightning on GitHub\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://arxiv.org/abs/2508.03680\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EAgent Lightning: Train ANY AI Agents with Reinforcement Learning\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://microsoft.github.io/agent-lightning\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EAgent Lightning documentation\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading {\"className\":\"h3\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EAgent Lightning: One learning system that makes all agents evolve\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELearning from experience is something humans do naturally, but for AI agents, it's often bolted on as an afterthought. Joining us from MSR Asia in Shanghai, Luna will introduce Agent Lightning, an open source framework that makes learning a first class capability for agents.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWith minimal code changes. Agents can improve over time using reinforcement learning and related methods. This project has taken off quickly in the community and is a great example of MSR research scaling through open source and shaping how people build a genetic systems in practice. Luna, over to you.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHello from Shanghai. I'm Luna Qiu, technical program manager of Microsoft Research Asia. And today let me introduce Agent Lightning, one learning system that makes all agents evolve. Our vision is straightforward. We want to build one system for every person and every organization to be able to evolve their own agents using their unique experience data and for individuals.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis means agents that genuinely understand your preferences, your needs. Through continuous learning. And for enterprises, that means distilling your specific workflows, your customers, your edge cases, into an intelligence layer. And that can be a compounding mode that competitors cannot replicate.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAgents exist on a wide spectrum. They vary in domains, implementation approaches, complexity, and more, while they all generate experience data. Turning that data into improvements can be extremely difficult. That's why at Microsoft Research, we open source agent lightning. To solve this, it connects any agent with any optimization method, making learning from experience practical with almost no code change.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe achieve this compatibility by going back to first principles, specifically Reinforcement Learning 101. Every RL setup can be formulated as a partially observable Markov decision process, where you can call it a POMDP. The RL agent takes actions, receive state and rewards from the RL environment without knowing all the details about how the environment works, and that is exactly how Agent Lightning goes.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIt abstracts such a POMDP from all AI executions. It treats only the part that needs to learn as scenario agent. Everything else becomes the RL environment. This design hides all the differences between AI agents inside the RL environment, so it is possible to optimize any agent using any learning method.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAnd there are three key benefits. One: a unified data pipeline that captures all agent activities using observability tools. It is non-intrusive to the original agent code. Two: algorithm flexibility that allows teams to plug in different methods without rebuilding everything. They can choose, from classical RL algorithms to novel methods that are tailored for agents.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis is important because the field is changing very fast and what the best approach is varies by use case. Three and finally, infrastructure disaggregation that separates agents and optimizers cleanly, providing modular components and a clear interface allowing independent scaling of each part as you need.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHere are some use cases from us and from the community showing how diverse the applications are. Let's start with the classic multi-agent customer service setting. So for this task, we used Agent Lightning to train a very small model. In just eight epochs, the agent with a 1.5 bit model achieves comparable performance to an agent using GPT 4 series, and we also enable a new algorithm called EMPO2.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIt is the first RL algorithm that can train agents with memory that significantly improves exploration with new out of distribution environments, and optimization is not limited to model parameters. One team at Microsoft is building a specialized coding agent for writing formal verification in rust, and with Agent Lightning prompt tuning at less than $6 per task, they improved the average success rate by over ten points.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAnd multimodal robotic agents also use it need to automatically tune their prompts. This improves the reasoning action coordination, doubling task success rate and significantly reduce completion time.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe community response has been extraordinary since open sourcing six months ago, Agent Lightning has earned more than 14,000 GitHub stars, ranking among Microsoft's top 50 most starred projects. It was featured as the number one project on GitHub and the trending research paper on Hugging Face.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis work reflects what we aspire to do at Microsoft Research. We tackle fundamental problems with open tools that serve the community and push the frontiers of what's possible. Our goal remains simple ,build one learning system that makes all agents evil, empowering every person and every organization to build intelligence that are truly their own.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAgent lightning is fully open source, so we welcome you to contribute and help shape what comes next. That's all I'm going to talk today. Thank you.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E\u003C/div\u003E\n\u003C!-- /wp:group --\u003E","post_title":"Agent Lightning: One learning system that makes all agents evolve","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"agent-lightning-one-learning-system-that-makes-all-agents-evolve","to_ping":"","pinged":"","post_modified":"2026-03-03 10:05:13","post_modified_gmt":"2026-03-03 18:05:13","post_content_filtered":"\n\u003Cdiv class=\"wp-block-group is-layout-constrained wp-block-group-is-layout-constrained\"\u003E\n\u003Cp\u003EAgent Lightning is an agent optimization framework that enables agents to learn from their experiences through reinforcement learning and other methods. By treating agents as first-class citizens, optimization becomes automatic for any agent with minimal code changes.\u003C/p\u003E\n\n\n\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://github.com/microsoft/agent-lightning\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EAgent Lightning on GitHub\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://arxiv.org/abs/2508.03680\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EAgent Lightning: Train ANY AI Agents with Reinforcement Learning\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://microsoft.github.io/agent-lightning\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EAgent Lightning documentation\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EAgent Lightning: One learning system that makes all agents evolve\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003ELearning from experience is something humans do naturally, but for AI agents, it’s often bolted on as an afterthought. Joining us from MSR Asia in Shanghai, Luna will introduce Agent Lightning, an open source framework that makes learning a first class capability for agents.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWith minimal code changes. Agents can improve over time using reinforcement learning and related methods. This project has taken off quickly in the community and is a great example of MSR research scaling through open source and shaping how people build a genetic systems in practice. Luna, over to you.\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-23\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003EHello from Shanghai. I’m Luna Qiu, technical program manager of Microsoft Research Asia. And today let me introduce Agent Lightning, one learning system that makes all agents evolve. Our vision is straightforward. We want to build one system for every person and every organization to be able to evolve their own agents using their unique experience data and for individuals.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis means agents that genuinely understand your preferences, your needs. Through continuous learning. And for enterprises, that means distilling your specific workflows, your customers, your edge cases, into an intelligence layer. And that can be a compounding mode that competitors cannot replicate.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAgents exist on a wide spectrum. They vary in domains, implementation approaches, complexity, and more, while they all generate experience data. Turning that data into improvements can be extremely difficult. That’s why at Microsoft Research, we open source agent lightning. To solve this, it connects any agent with any optimization method, making learning from experience practical with almost no code change.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe achieve this compatibility by going back to first principles, specifically Reinforcement Learning 101. Every RL setup can be formulated as a partially observable Markov decision process, where you can call it a POMDP. The RL agent takes actions, receive state and rewards from the RL environment without knowing all the details about how the environment works, and that is exactly how Agent Lightning goes.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIt abstracts such a POMDP from all AI executions. It treats only the part that needs to learn as scenario agent. Everything else becomes the RL environment. This design hides all the differences between AI agents inside the RL environment, so it is possible to optimize any agent using any learning method.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAnd there are three key benefits. One: a unified data pipeline that captures all agent activities using observability tools. It is non-intrusive to the original agent code. Two: algorithm flexibility that allows teams to plug in different methods without rebuilding everything. They can choose, from classical RL algorithms to novel methods that are tailored for agents.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis is important because the field is changing very fast and what the best approach is varies by use case. Three and finally, infrastructure disaggregation that separates agents and optimizers cleanly, providing modular components and a clear interface allowing independent scaling of each part as you need.\u003C/p\u003E\n\n\n\n\u003Cp\u003EHere are some use cases from us and from the community showing how diverse the applications are. Let’s start with the classic multi-agent customer service setting. So for this task, we used Agent Lightning to train a very small model. In just eight epochs, the agent with a 1.5 bit model achieves comparable performance to an agent using GPT 4 series, and we also enable a new algorithm called EMPO2.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIt is the first RL algorithm that can train agents with memory that significantly improves exploration with new out of distribution environments, and optimization is not limited to model parameters. One team at Microsoft is building a specialized coding agent for writing formal verification in rust, and with Agent Lightning prompt tuning at less than $6 per task, they improved the average success rate by over ten points.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAnd multimodal robotic agents also use it need to automatically tune their prompts. This improves the reasoning action coordination, doubling task success rate and significantly reduce completion time.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe community response has been extraordinary since open sourcing six months ago, Agent Lightning has earned more than 14,000 GitHub stars, ranking among Microsoft’s top 50 most starred projects. It was featured as the number one project on GitHub and the trending research paper on Hugging Face.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis work reflects what we aspire to do at Microsoft Research. We tackle fundamental problems with open tools that serve the community and push the frontiers of what’s possible. Our goal remains simple ,build one learning system that makes all agents evil, empowering every person and every organization to build intelligence that are truly their own.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAgent lightning is fully open source, so we welcome you to contribute and help shape what comes next. That’s all I’m going to talk today. Thank you.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-23\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/agent-lightning-one-learning-system-that-makes-all-agents-evolve/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-post-option":[{"term_id":269148,"slug":"approved-for-river","name":"Approved for River","parent":0,"term_taxonomy_id":269193,"term_order":0,"facet":"{\"term_id\":269148,\"slug\":\"approved-for-river\",\"name\":\"Approved for River\",\"parent\":0,\"term_taxonomy_id\":269193,\"term_order\":0}"},{"term_id":269142,"slug":"include-in-river","name":"Include in River","parent":0,"term_taxonomy_id":269187,"term_order":0,"facet":"{\"term_id\":269142,\"slug\":\"include-in-river\",\"name\":\"Include in River\",\"parent\":0,\"term_taxonomy_id\":269187,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270235,"slug":"season-2-episode-3","name":"Season 2, Episode 3","parent":0,"term_taxonomy_id":270280,"term_order":0,"facet":"{\"term_id\":270235,\"slug\":\"season-2-episode-3\",\"name\":\"Season 2, Episode 3\",\"parent\":0,\"term_taxonomy_id\":270280,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270108,"slug":"impact-at-scale","name":"Impact at scale","parent":0,"term_taxonomy_id":270153,"term_order":0,"facet":"{\"term_id\":270108,\"slug\":\"impact-at-scale\",\"name\":\"Impact at scale\",\"parent\":0,\"term_taxonomy_id\":270153,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"228\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Luna_01-480x280.png\" class=\"card-img wp-post-image\" alt=\"photo of Luna Qiu during the Microsoft Research Forum\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t06:30\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Luna_01-480x280.png\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/agent-lightning-one-learning-system-that-makes-all-agents-evolve/\" data-bi-cN=\"Agent Lightning: One learning system that makes all agents evolve\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled Agent Lightning: One learning system that makes all agents evolve\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003EAgent Lightning: One learning system that makes all agents evolve\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tMarch 3, 2026\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/ansonho/\"\u003EAnson Ho\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/lunaqiu/\"\u003ELuna K. Qiu\u003C/a\u003E\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 3\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"},{"data":{"ID":1162537,"post_author":43868,"post_date":"2026-03-03 10:05:09","post_date_gmt":"2026-03-03 18:05:09","post_content":"\u003C!-- wp:group {\"layout\":{\"type\":\"constrained\"}} --\u003E\n\u003Cdiv class=\"wp-block-group\"\u003E\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAs AI agents move from isolated tools to active participants in multi-agent ecosystems, their success depends on more than task competence\u2014it requires strategic behavior under misaligned incentives and imperfect information. Using Magentic Marketplace, an open-source simulation of two-sided agent markets, we show that while frontier models can achieve strong welfare outcomes in ideal settings, performance degrades at scale and reveals emergent failure modes such as manipulation and speed bias, motivating a shift toward training agents for social reasoning.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"className\":\"h5\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/magentic-marketplace-an-open-source-environment-for-studying-agentic-markets/\"\u003EMagentic Marketplace: An Open-Source Environment for Studying Agentic Markets\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://github.com/microsoft/multi-agent-marketplace\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EMagentic Marketplace on GitHub\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/blog/magentic-marketplace-an-open-source-simulation-environment-for-studying-agentic-markets/\"\u003EMagentic Marketplace: an open-source simulation environment for studying agentic markets\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading {\"className\":\"h3\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EMagentic Marketplace: Testing societies of agents at scale\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAs AI agents become more capable, they are no longer working alone. Agents start interacting with each other in shared environments. And when that happens, things can get interesting here in Redmond. Gagan Bansal from MSR, AI frontiers studies what happens when agents start transacting in a market-like settings using an open source environment called Magentic Marketplace.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHis team explores how agents behave when incentives don't personally line up, and what kind of failure modes start to appear as systems scale.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis is exactly the kind of early curiosity driven research that helps us understand what could go wrong before these systems light up. Take it away.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EHi, I'm Gagan Bansal and I'm a researcher at Microsoft. And today I want to talk to you about our recent work on applying societies of agents in markets. And although I'm presenting, this work was a collaboration between many amazing colleagues across Microsoft.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ECapabilities of AI agents are improving rapidly. We're quickly moving towards a future where each one of us will have personal agents. Now, in a world where everyone has agents, we believe societies of agents will drive new applications where our agents will have to interact with other agents. But how can we trust agents that we don't control, or agents that might know things ours doesn't, or even have competing goals?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAt Microsoft, we've been building on our expertise in multi-agent frameworks like AutoGen and Magentic one to create useful societies of agents, ones that add value, save time, and don't cause harm. To enable this future, we need to understand how agents behave when they interact at scale. Recent examples from the open source community, where agents could talk freely on the forums, only underscores how timely and important this question is.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELet me show you what we built and what we found. Imagine a marketplace where all the buying and selling is done by agents representing people. We call these settings as two sided agentic markets. This setting is a great testbed for society's agents, because every agent has access to different information and competing incentives to systematically study two sided markets.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe at Microsoft Research built a new simulation environment called Magentic Marketplace. Here, assistant agents represent customers, service agents represents businesses, and a marketplace sits in the middle handles search, communication and transaction between agents. Here's a typical interaction.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESuppose a customer wants to find a restaurant with something specific, like delicious empanadas and outdoor seating. Their assistant can search the marketplace. Talk to the service agents. Ask about menus, check amenities, and finally make a reservation. This framework allows us to test hundreds of agents buying and selling in parallel.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe used it to systematically ask many research questions. Do these agents even add value for customers and businesses? Does the quality of search result impact their behavior? Are they vulnerable to any biases or manipulation? We built this framework as a general research tool. It can be used to ask many other questions, even for domains beyond markets.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe started by asking whether agents actually add value for consumers. To find out, we implemented agents using frontier and open source models and computed the welfare that they achieve. Here, welfare is the value customers get from their purchase minus the price they paid. Higher is better. We observed that when agents have access to high quality search results, frontier models like GPT 5 and Sonnet 4 reached near optimal welfare.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThey talk to business agents, gathered missing information and made good choices. But we found that agent performance was tied to the quality of search results. When the search quality dropped, performance dropped. We also observed that there is still a large gap between the welfare achieved by frontier models and open source models.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn addition to the impact of the quality of search results, we also want to test whether the number of search results impact welfare. So we gave agents more search results, varying them from 3 to 100, and expected the welfare to increase. But the opposite happened, resulting in a surprising paradox of choice.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWelfare dropped for almost every model. This happened because agents didn't explore enough and contacted few businesses. We also conducted experiments that tested whether the order of offers from service agents matters. It did dramatically. Almost 80 to 100% of the agents accepted the first proposal they received.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThey never even looked at the alternatives. Think about what this means for a real market. Speed beats quality. A business gains more from responding fast than from offering a better deal. That's not a healthy dynamic. We also tested vulnerability to fake reviews, fake awards and prompt injection.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ESome frontier models resisted everything, but others were completely compromised. All payments were redirected to the attackers. These are early findings and markets are just the beginning. Societies of agents will emerge anywhere. Agents represent people with different interests such as supply chains, hiring and negotiation.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EMagnetic marketplace is open source and GitHub for the community to run experiments, stress test agents and help answer the harder questions. What guardrails do we need? How should markets be designed when both sides are AI? What we've shown is that simulation matters. Agents can add value, but they also inherit biases, fall for manipulation, and make choices that reward speed over quality.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThese are not edge cases. These are behaviors that only emerge when societies of agents are tested at scale. If these agents are going to take high stakes decisions on our behalf, such as making transactions to other agents, we should understand their behaviors and biases before deployment and not after.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EPlease check out our papers and GitHub repository for more information. And thank you for attending the Microsoft Research Forum.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E\u003C/div\u003E\n\u003C!-- /wp:group --\u003E","post_title":"Magentic Marketplace: Testing societies of agents at scale","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"magentic-marketplace-testing-societies-of-agents-at-scale","to_ping":"","pinged":"","post_modified":"2026-03-03 10:05:10","post_modified_gmt":"2026-03-03 18:05:10","post_content_filtered":"\n\u003Cdiv class=\"wp-block-group is-layout-constrained wp-block-group-is-layout-constrained\"\u003E\n\u003Cp\u003EAs AI agents move from isolated tools to active participants in multi-agent ecosystems, their success depends on more than task competence\u2014it requires strategic behavior under misaligned incentives and imperfect information. Using Magentic Marketplace, an open-source simulation of two-sided agent markets, we show that while frontier models can achieve strong welfare outcomes in ideal settings, performance degrades at scale and reveals emergent failure modes such as manipulation and speed bias, motivating a shift toward training agents for social reasoning.\u003C/p\u003E\n\n\n\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/magentic-marketplace-an-open-source-environment-for-studying-agentic-markets/\"\u003EMagentic Marketplace: An Open-Source Environment for Studying Agentic Markets\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://github.com/microsoft/multi-agent-marketplace\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EMagentic Marketplace on GitHub\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/blog/magentic-marketplace-an-open-source-simulation-environment-for-studying-agentic-markets/\"\u003EMagentic Marketplace: an open-source simulation environment for studying agentic markets\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EMagentic Marketplace: Testing societies of agents at scale\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003EAs AI agents become more capable, they are no longer working alone. Agents start interacting with each other in shared environments. And when that happens, things can get interesting here in Redmond. Gagan Bansal from MSR, AI frontiers studies what happens when agents start transacting in a market-like settings using an open source environment called Magentic Marketplace.\u003C/p\u003E\n\n\n\n\u003Cp\u003EHis team explores how agents behave when incentives don’t personally line up, and what kind of failure modes start to appear as systems scale.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis is exactly the kind of early curiosity driven research that helps us understand what could go wrong before these systems light up. Take it away.\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-24\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003EHi, I’m Gagan Bansal and I’m a researcher at Microsoft. And today I want to talk to you about our recent work on applying societies of agents in markets. And although I’m presenting, this work was a collaboration between many amazing colleagues across Microsoft.\u003C/p\u003E\n\n\n\n\u003Cp\u003ECapabilities of AI agents are improving rapidly. We’re quickly moving towards a future where each one of us will have personal agents. Now, in a world where everyone has agents, we believe societies of agents will drive new applications where our agents will have to interact with other agents. But how can we trust agents that we don’t control, or agents that might know things ours doesn’t, or even have competing goals?\u003C/p\u003E\n\n\n\n\u003Cp\u003EAt Microsoft, we’ve been building on our expertise in multi-agent frameworks like AutoGen and Magentic one to create useful societies of agents, ones that add value, save time, and don’t cause harm. To enable this future, we need to understand how agents behave when they interact at scale. Recent examples from the open source community, where agents could talk freely on the forums, only underscores how timely and important this question is.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELet me show you what we built and what we found. Imagine a marketplace where all the buying and selling is done by agents representing people. We call these settings as two sided agentic markets. This setting is a great testbed for society’s agents, because every agent has access to different information and competing incentives to systematically study two sided markets.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe at Microsoft Research built a new simulation environment called Magentic Marketplace. Here, assistant agents represent customers, service agents represents businesses, and a marketplace sits in the middle handles search, communication and transaction between agents. Here’s a typical interaction.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESuppose a customer wants to find a restaurant with something specific, like delicious empanadas and outdoor seating. Their assistant can search the marketplace. Talk to the service agents. Ask about menus, check amenities, and finally make a reservation. This framework allows us to test hundreds of agents buying and selling in parallel.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe used it to systematically ask many research questions. Do these agents even add value for customers and businesses? Does the quality of search result impact their behavior? Are they vulnerable to any biases or manipulation? We built this framework as a general research tool. It can be used to ask many other questions, even for domains beyond markets.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe started by asking whether agents actually add value for consumers. To find out, we implemented agents using frontier and open source models and computed the welfare that they achieve. Here, welfare is the value customers get from their purchase minus the price they paid. Higher is better. We observed that when agents have access to high quality search results, frontier models like GPT 5 and Sonnet 4 reached near optimal welfare.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThey talk to business agents, gathered missing information and made good choices. But we found that agent performance was tied to the quality of search results. When the search quality dropped, performance dropped. We also observed that there is still a large gap between the welfare achieved by frontier models and open source models.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn addition to the impact of the quality of search results, we also want to test whether the number of search results impact welfare. So we gave agents more search results, varying them from 3 to 100, and expected the welfare to increase. But the opposite happened, resulting in a surprising paradox of choice.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWelfare dropped for almost every model. This happened because agents didn’t explore enough and contacted few businesses. We also conducted experiments that tested whether the order of offers from service agents matters. It did dramatically. Almost 80 to 100% of the agents accepted the first proposal they received.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThey never even looked at the alternatives. Think about what this means for a real market. Speed beats quality. A business gains more from responding fast than from offering a better deal. That’s not a healthy dynamic. We also tested vulnerability to fake reviews, fake awards and prompt injection.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESome frontier models resisted everything, but others were completely compromised. All payments were redirected to the attackers. These are early findings and markets are just the beginning. Societies of agents will emerge anywhere. Agents represent people with different interests such as supply chains, hiring and negotiation.\u003C/p\u003E\n\n\n\n\u003Cp\u003EMagnetic marketplace is open source and GitHub for the community to run experiments, stress test agents and help answer the harder questions. What guardrails do we need? How should markets be designed when both sides are AI? What we’ve shown is that simulation matters. Agents can add value, but they also inherit biases, fall for manipulation, and make choices that reward speed over quality.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThese are not edge cases. These are behaviors that only emerge when societies of agents are tested at scale. If these agents are going to take high stakes decisions on our behalf, such as making transactions to other agents, we should understand their behaviors and biases before deployment and not after.\u003C/p\u003E\n\n\n\n\u003Cp\u003EPlease check out our papers and GitHub repository for more information. And thank you for attending the Microsoft Research Forum.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-24\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/magentic-marketplace-testing-societies-of-agents-at-scale/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-post-option":[{"term_id":269148,"slug":"approved-for-river","name":"Approved for River","parent":0,"term_taxonomy_id":269193,"term_order":0,"facet":"{\"term_id\":269148,\"slug\":\"approved-for-river\",\"name\":\"Approved for River\",\"parent\":0,\"term_taxonomy_id\":269193,\"term_order\":0}"},{"term_id":269142,"slug":"include-in-river","name":"Include in River","parent":0,"term_taxonomy_id":269187,"term_order":0,"facet":"{\"term_id\":269142,\"slug\":\"include-in-river\",\"name\":\"Include in River\",\"parent\":0,\"term_taxonomy_id\":269187,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270235,"slug":"season-2-episode-3","name":"Season 2, Episode 3","parent":0,"term_taxonomy_id":270280,"term_order":0,"facet":"{\"term_id\":270235,\"slug\":\"season-2-episode-3\",\"name\":\"Season 2, Episode 3\",\"parent\":0,\"term_taxonomy_id\":270280,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270112,"slug":"research-exploration","name":"Open research exploration","parent":0,"term_taxonomy_id":270157,"term_order":0,"facet":"{\"term_id\":270112,\"slug\":\"research-exploration\",\"name\":\"Open research exploration\",\"parent\":0,\"term_taxonomy_id\":270157,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"206\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Gagan_01-scaled.png\" class=\"card-img wp-post-image\" alt=\"photo of Gagan Basal during the Microsoft Research Forum\" srcset=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Gagan_01-scaled.png 2560w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Gagan_01-300x158.png 300w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Gagan_01-1024x540.png 1024w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Gagan_01-768x405.png 768w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Gagan_01-1536x810.png 1536w, https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Gagan_01-2048x1080.png 2048w\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t06:20\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2026/02/Gagan_01-scaled.png\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/magentic-marketplace-testing-societies-of-agents-at-scale/\" data-bi-cN=\"Magentic Marketplace: Testing societies of agents at scale\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled Magentic Marketplace: Testing societies of agents at scale\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003EMagentic Marketplace: Testing societies of agents at scale\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tMarch 3, 2026\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/gaganbansal/\"\u003EGagan Bansal\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/ansonho/\"\u003EAnson Ho\u003C/a\u003E\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 3\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"},{"data":{"ID":1157621,"post_author":43868,"post_date":"2025-12-09 08:26:00","post_date_gmt":"2025-12-09 16:26:00","post_content":"\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EData Formulator is an AI-powered tool for analysts to iteratively explore and visualize data. Started with data in any format (screenshot, text, csv, or database), users can work with AI agents with a novel blended interface that combines user interface interactions (UI) and natural language (NL) inputs to communicate their intents, control branching exploration directions, and create reports to share their insights.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:heading {\"className\":\"h5\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:list --\u003E\n\u003Cul class=\"wp-block-list\"\u003E\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://github.com/microsoft/data-formulator\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EData Formulator on GitHub\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/data-formulator-2-iteratively-creating-rich-visualizations-with-ai/\"\u003EData Formulator 2: Iterative Creation of Data Visualizations, with AI Transforming Data Along the Way\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\n\n\u003C!-- wp:list-item --\u003E\n\u003Cli\u003E\u003Ca href=\"https://labs.ai.azure.com/projects/data-formulator/\" target=\"_blank\" rel=\"noreferrer noopener\"\u003EData Formulator on Azure AI Foundry Labs\u003C/a\u003E\u003C/li\u003E\n\u003C!-- /wp:list-item --\u003E\u003C/ul\u003E\n\u003C!-- /wp:list --\u003E\n\n\u003C!-- wp:buttons --\u003E\n\u003Cdiv class=\"wp-block-buttons\"\u003E\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\n\n\u003C!-- wp:button {\"className\":\"is-style-cta\"} --\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C!-- /wp:button --\u003E\u003C/div\u003E\n\u003C!-- /wp:buttons --\u003E\n\n\u003C!-- wp:msr/show-more --\u003E\n\u003C!-- wp:heading {\"className\":\"h3\"} --\u003E\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\u003C!-- /wp:heading --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EData Formulator: Vibe with your data, in control\u003C/strong\u003E\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph {\"placeholder\":\"More content...\"} --\u003E\n\u003Cp\u003E[MUSIC]\u202f \u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO SWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003EKAREN EASTERBROOK: \u003C/strong\u003EAt Microsoft Research, we see fresh innovations emerge out of the labs ready to reshape how people and organizations harness technology in their everyday work.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EA fellow Redmond-based colleague, Chenglong Wang, principal researcher, works with his team to reimagine how analysts interact with data. By combining natural language and intuitive interface design, they introduced Data Formulator. This tool helps users guide AI collaboratively\u2014to visualize, explore, and truly \u201cvibe\u201d with data, in control.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EOver to you, Chenglong.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:msr/more --\u003E\n\u003C!--msr/more--\u003E\n\u003C!-- /wp:msr/more --\u003E\n\n\u003C!-- wp:paragraph {\"placeholder\":\"More content...\"} --\u003E\n\u003Cp\u003E[MUSIC]\u202f \u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E[MUSIC FADES INTO SWEEPING SOUND]\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003E\u003Cstrong\u003ECHENGLONG WANG: \u003C/strong\u003EHello, everyone. I'm Chenglong from Microsoft Research Redmond. Today, I'd like to introduce you [to] Data Formulator. It's an interactive AI-powered visualization tool for you to explore data with AI agents, in control.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWhy do we build this tool? For example, given a dataset about US product prices from 2005 to 2025, we may be curious about, how much does a product price grow over the years, as everything around us has become more expensive lately, right? The best way to answer this question is to explore the data and show insights with some visualizations. But how can we explore effectively?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ETo get these insights, we need to go through an iterative data exploration process.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe first need to formulate plans based on our data. Then we need programming skills to transform and implement the visualizations. We often don't solve it in one step. We may need to go back to take another branch, to backtrack, and revisit some steps. After a few iterations, we should find some insights and we're ready to share.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis process, however, can be challenging. It requires a user to have both data science expertise to formulate good questions and also requires good programming skills to implement these designs. How can we make this accessible to everyone who is interested in data?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELuckily, with AI agents that can generate code from user instructions, coding is no longer a big barrier. However, to make data analysis truly accessible, we still need to lower the interaction friction between user and AI.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThe first challenge is that natural language can be quite universal, but it can be verbose for describing the visualization intent and it may not be very precise. The second is that chat interface is a way how we go with interact with agents, but it doesn't support iteration naturally as we want for data analysis. Third, agents can be difficult to control if they are working in a black box. The user may easily lose control over what agent is doing.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIn order to solve this problem, we introduce Data Formulator. It's an interactive data analysis tool for people to work with agents effectively to explore data.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIt first features a multi-modal interaction. The multi-modal UI combines UI interaction and natural language so that users can specify their design both precisely and concisely. Second, to support iteration, we organize exploration into threads, allowing users to backtrack, follow up, and revisit. On top of that, we developed agent mode. Instead of running in a black box, it operates on top of the data thread so the user can take control whenever they want.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELet me show you the experience of using Data Formulator to explore data.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ELet's explore, how much do US product prices increase over the decades? Let's first load the data. \u003Cem\u003EThis\u003C/em\u003E data. We can start with agent mode; ask AI agent to automatically explore the data on our behalf. The agent first generates a plan, and then it's trying to generate code to execute the plan by transforming the data and visualize it.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EThis is the first chart recommended by the agent. It's a line chart for temporal price trends. Wow, everything has been more expensive now.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIt's now recommended a second chart. It's a bar chart to show the product volatility. It seems egg price is most volatile.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ENow we can take control; we can ask a follow-up question using natural language: how does COVID and 2008 crisis affect prices?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EIt now generates a stacked bar chart, not certainly the best way to visualize it. We can use UI to change it to a grouped bar chart. Now it's clear. It seems that COVID affected prices more significantly. That's very interesting.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe can now go back to the previous branch to explore something different. Hmm, what should we explore here? We can ask the AI agent for some ideas. It seems it's interesting to see the price correlation with gas price. Let's go with that.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003ENow we get a bar chart to show the price correlation with gas price. It seems that most product prices are correlated to the gas, except \u2026 tomato?\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EWe can continue to explore, but now we can also write a report to share our findings. We can ask AI agents to generate a report grounded on our data. It will leverage the computation, data, and charts to compose a report.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph --\u003E\n\u003Cp\u003EAs you can see, with AI agents automating the implementations, we can easily vibe with our data. The blended UI-plus-natural-language-interaction approach allows us to easily work with agents. While agent is automating both steps, we can easily take control when we need to go to the exploration path we like.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\n\u003C!-- wp:paragraph {\"placeholder\":\"More content...\"} --\u003E\n\u003Cp\u003ESo this is Data Formulator. It is an interactive tool for people to explore and visualize data. It features a multi-modal interface that makes it easy for users to specify visualization ideas. It's providing a new design, data threads, so that users can perform nonlinear interactions with AI agents in exploration tasks. Thirdly, the agent plus interactive modes allow users to vibe with data but still having full control. We invite you to try Data Formulator on data-formulator.ai. Explore with some data and let us know what you find.\u003C/p\u003E\n\u003C!-- /wp:paragraph --\u003E\n\u003C!-- /wp:msr/show-more --\u003E","post_title":"Data Formulator: Vibe with your data, in control","post_excerpt":"","post_status":"publish","comment_status":"closed","ping_status":"closed","post_password":"","post_name":"data-formulator-vibe-with-your-data-in-control","to_ping":"","pinged":"","post_modified":"2025-12-09 13:18:53","post_modified_gmt":"2025-12-09 21:18:53","post_content_filtered":"\n\u003Cp\u003EData Formulator is an AI-powered tool for analysts to iteratively explore and visualize data. Started with data in any format (screenshot, text, csv, or database), users can work with AI agents with a novel blended interface that combines user interface interactions (UI) and natural language (NL) inputs to communicate their intents, control branching exploration directions, and create reports to share their insights.\u003C/p\u003E\n\n\n\n\u003Ch2 class=\"wp-block-heading h5\" id=\"explore-more\"\u003EExplore more\u003C/h2\u003E\n\n\n\n\u003Cul class=\"wp-block-list\"\u003E\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://github.com/microsoft/data-formulator\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EData Formulator on GitHub\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca href=\"https://www.microsoft.com/en-us/research/publication/data-formulator-2-iteratively-creating-rich-visualizations-with-ai/\"\u003EData Formulator 2: Iterative Creation of Data Visualizations, with AI Transforming Data Along the Way\u003C/a\u003E\u003C/li\u003E\n\n\n\n\u003Cli\u003E\u003Ca class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https://labs.ai.azure.com/projects/data-formulator/\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EData Formulator on Azure AI Foundry Labs\u003Cspan class=\"sr-only\"\u003E (opens in new tab)\u003C/span\u003E\u003C/a\u003E\u003C/li\u003E\n\u003C/ul\u003E\n\n\n\n\u003Cdiv class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\"\u003E\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"https://www.microsoft.com/en-us/research/event/microsoft-research-forum/past-episodes/\"\u003EAll Research Forum sessions\u003C/a\u003E\u003C/div\u003E\n\n\n\n\u003Cdiv class=\"wp-block-button is-style-cta\"\u003E\u003Ca data-bi-type=\"button\" class=\"wp-block-button__link wp-element-button\" href=\"http://aka.ms/researchforum-register\" target=\"_blank\" rel=\"noreferrer noopener\"\u003ERegister for the series\u003C/a\u003E\u003C/div\u003E\n\u003C/div\u003E\n\n\n\u003Cdiv class=\"wp-block-msr-show-more\"\u003E\n\t\u003Cdiv class=\"bg-neutral-100 p-5\"\u003E\n\t\t\u003Cdiv class=\"show-more-show-less\"\u003E\n\t\t\t\u003Cdiv\u003E\n\t\t\t\t\u003Cspan\u003E\n\t\t\t\t\t\n\n\u003Ch2 class=\"wp-block-heading h3\" id=\"transcript\"\u003ETranscript\u003C/h2\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EData Formulator: Vibe with your data, in control\u003C/strong\u003E\u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC]\u202f \u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO SWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003EKAREN EASTERBROOK: \u003C/strong\u003EAt Microsoft Research, we see fresh innovations emerge out of the labs ready to reshape how people and organizations harness technology in their everyday work.\u003C/p\u003E\n\n\n\n\u003Cp\u003EA fellow Redmond-based colleague, Chenglong Wang, principal researcher, works with his team to reimagine how analysts interact with data. By combining natural language and intuitive interface design, they introduced Data Formulator. This tool helps users guide AI collaboratively\u2014to visualize, explore, and truly \u201cvibe\u201d with data, in control.\u003C/p\u003E\n\n\n\n\u003Cp\u003EOver to you, Chenglong.\u003C/p\u003E\n\n\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\t\u003Cspan id=\"show-more-show-less-toggle-41\" class=\"show-more-show-less-toggleable-content\"\u003E\n\t\t\t\t\t\n\n\n\n\u003Cp\u003E[MUSIC]\u202f \u003C/p\u003E\n\n\n\n\u003Cp\u003E[MUSIC FADES INTO SWEEPING SOUND]\u003C/p\u003E\n\n\n\n\u003Cp\u003E\u003Cstrong\u003ECHENGLONG WANG: \u003C/strong\u003EHello, everyone. I’m Chenglong from Microsoft Research Redmond. Today, I’d like to introduce you [to] Data Formulator. It’s an interactive AI-powered visualization tool for you to explore data with AI agents, in control.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWhy do we build this tool? For example, given a dataset about US product prices from 2005 to 2025, we may be curious about, how much does a product price grow over the years, as everything around us has become more expensive lately, right? The best way to answer this question is to explore the data and show insights with some visualizations. But how can we explore effectively?\u003C/p\u003E\n\n\n\n\u003Cp\u003ETo get these insights, we need to go through an iterative data exploration process.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe first need to formulate plans based on our data. Then we need programming skills to transform and implement the visualizations. We often don’t solve it in one step. We may need to go back to take another branch, to backtrack, and revisit some steps. After a few iterations, we should find some insights and we’re ready to share.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis process, however, can be challenging. It requires a user to have both data science expertise to formulate good questions and also requires good programming skills to implement these designs. How can we make this accessible to everyone who is interested in data?\u003C/p\u003E\n\n\n\n\u003Cp\u003ELuckily, with AI agents that can generate code from user instructions, coding is no longer a big barrier. However, to make data analysis truly accessible, we still need to lower the interaction friction between user and AI.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThe first challenge is that natural language can be quite universal, but it can be verbose for describing the visualization intent and it may not be very precise. The second is that chat interface is a way how we go with interact with agents, but it doesn’t support iteration naturally as we want for data analysis. Third, agents can be difficult to control if they are working in a black box. The user may easily lose control over what agent is doing.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIn order to solve this problem, we introduce Data Formulator. It’s an interactive data analysis tool for people to work with agents effectively to explore data.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIt first features a multi-modal interaction. The multi-modal UI combines UI interaction and natural language so that users can specify their design both precisely and concisely. Second, to support iteration, we organize exploration into threads, allowing users to backtrack, follow up, and revisit. On top of that, we developed agent mode. Instead of running in a black box, it operates on top of the data thread so the user can take control whenever they want.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELet me show you the experience of using Data Formulator to explore data.\u003C/p\u003E\n\n\n\n\u003Cp\u003ELet’s explore, how much do US product prices increase over the decades? Let’s first load the data. \u003Cem\u003EThis\u003C/em\u003E data. We can start with agent mode; ask AI agent to automatically explore the data on our behalf. The agent first generates a plan, and then it’s trying to generate code to execute the plan by transforming the data and visualize it.\u003C/p\u003E\n\n\n\n\u003Cp\u003EThis is the first chart recommended by the agent. It’s a line chart for temporal price trends. Wow, everything has been more expensive now.\u003C/p\u003E\n\n\n\n\u003Cp\u003EIt’s now recommended a second chart. It’s a bar chart to show the product volatility. It seems egg price is most volatile.\u003C/p\u003E\n\n\n\n\u003Cp\u003ENow we can take control; we can ask a follow-up question using natural language: how does COVID and 2008 crisis affect prices?\u003C/p\u003E\n\n\n\n\u003Cp\u003EIt now generates a stacked bar chart, not certainly the best way to visualize it. We can use UI to change it to a grouped bar chart. Now it’s clear. It seems that COVID affected prices more significantly. That’s very interesting.\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe can now go back to the previous branch to explore something different. Hmm, what should we explore here? We can ask the AI agent for some ideas. It seems it’s interesting to see the price correlation with gas price. Let’s go with that.\u003C/p\u003E\n\n\n\n\u003Cp\u003ENow we get a bar chart to show the price correlation with gas price. It seems that most product prices are correlated to the gas, except \u2026 tomato?\u003C/p\u003E\n\n\n\n\u003Cp\u003EWe can continue to explore, but now we can also write a report to share our findings. We can ask AI agents to generate a report grounded on our data. It will leverage the computation, data, and charts to compose a report.\u003C/p\u003E\n\n\n\n\u003Cp\u003EAs you can see, with AI agents automating the implementations, we can easily vibe with our data. The blended UI-plus-natural-language-interaction approach allows us to easily work with agents. While agent is automating both steps, we can easily take control when we need to go to the exploration path we like.\u003C/p\u003E\n\n\n\n\u003Cp\u003ESo this is Data Formulator. It is an interactive tool for people to explore and visualize data. It features a multi-modal interface that makes it easy for users to specify visualization ideas. It’s providing a new design, data threads, so that users can perform nonlinear interactions with AI agents in exploration tasks. Thirdly, the agent plus interactive modes allow users to vibe with data but still having full control. We invite you to try Data Formulator on data-formulator.ai. Explore with some data and let us know what you find.\u003C/p\u003E\n\n\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\u003Cbutton\n\t\t\t\tclass=\"action-trigger glyph-prepend mt-2 mb-0 show-more-show-less-toggle\"\n\t\t\t\taria-expanded=\"false\"\n\t\t\t\tdata-show-less-text=\"Show less\"\n\t\t\t\ttype=\"button\"\n\t\t\t\taria-controls=\"show-more-show-less-toggle-41\"\n\t\t\t\taria-label=\"Show more content\"\n\t\t\t\tdata-alternate-aria-label=\"Show less content\"\u003E\n\t\t\t\tShow more\t\t\t\u003C/button\u003E\n\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n\u003Cspan id=\"label-external-link\" class=\"sr-only\" aria-hidden=\"true\"\u003EOpens in a new tab\u003C/span\u003E","post_parent":0,"guid":"","menu_order":0,"post_type":"msr-video","post_mime_type":"","comment_count":0,"filter":"raw","site_id":1,"permalink":"https://www.microsoft.com/en-us/research/video/data-formulator-vibe-with-your-data-in-control/","terms":{"msr-research-area":[{"term_id":13556,"slug":"artificial-intelligence","name":"Artificial intelligence","parent":0,"term_taxonomy_id":13556,"term_order":0,"facet":"{\"term_id\":13556,\"slug\":\"artificial-intelligence\",\"name\":\"Artificial intelligence\",\"parent\":0,\"term_taxonomy_id\":13556,\"term_order\":0}"},{"term_id":13554,"slug":"human-computer-interaction","name":"Human-computer interaction","parent":0,"term_taxonomy_id":13554,"term_order":0,"facet":"{\"term_id\":13554,\"slug\":\"human-computer-interaction\",\"name\":\"Human-computer interaction\",\"parent\":0,\"term_taxonomy_id\":13554,\"term_order\":0}"}],"msr-video-type":[{"term_id":268311,"slug":"microsoft-research-forum","name":"Microsoft Research Forum","parent":0,"term_taxonomy_id":268356,"term_order":0,"facet":"{\"term_id\":268311,\"slug\":\"microsoft-research-forum\",\"name\":\"Microsoft Research Forum\",\"parent\":0,\"term_taxonomy_id\":268356,\"term_order\":0}"}],"msr-locale":[{"term_id":268875,"slug":"en_us","name":"English","parent":0,"term_taxonomy_id":268920,"term_order":0,"facet":"{\"term_id\":268875,\"slug\":\"en_us\",\"name\":\"English\",\"parent\":0,\"term_taxonomy_id\":268920,\"term_order\":0}"}],"msr-post-option":[{"term_id":269148,"slug":"approved-for-river","name":"Approved for River","parent":0,"term_taxonomy_id":269193,"term_order":0,"facet":"{\"term_id\":269148,\"slug\":\"approved-for-river\",\"name\":\"Approved for River\",\"parent\":0,\"term_taxonomy_id\":269193,\"term_order\":0}"},{"term_id":269142,"slug":"include-in-river","name":"Include in River","parent":0,"term_taxonomy_id":269187,"term_order":0,"facet":"{\"term_id\":269142,\"slug\":\"include-in-river\",\"name\":\"Include in River\",\"parent\":0,\"term_taxonomy_id\":269187,\"term_order\":0}"}],"msr-session-type":[{"term_id":256174,"slug":"talk","name":"Talk","parent":0,"term_taxonomy_id":256201,"term_order":0,"facet":"{\"term_id\":256174,\"slug\":\"talk\",\"name\":\"Talk\",\"parent\":0,\"term_taxonomy_id\":256201,\"term_order\":0}"}],"msr-episode":[{"term_id":270184,"slug":"season-2-episode-2","name":"Season 2, Episode 2","parent":0,"term_taxonomy_id":270229,"term_order":0,"facet":"{\"term_id\":270184,\"slug\":\"season-2-episode-2\",\"name\":\"Season 2, Episode 2\",\"parent\":0,\"term_taxonomy_id\":270229,\"term_order\":0}"}],"msr-research-theme":[{"term_id":270110,"slug":"new-advancements-out-of-the-lab","name":"New advancement out of the lab","parent":0,"term_taxonomy_id":270155,"term_order":0,"facet":"{\"term_id\":270110,\"slug\":\"new-advancements-out-of-the-lab\",\"name\":\"New advancement out of the lab\",\"parent\":0,\"term_taxonomy_id\":270155,\"term_order\":0}"}]},"meta":[],"elasticsearch":true},"markup":"\n\u003C!-- Card column wrapper --\u003E\n\u003Cdiv class=\"col mb-4\" itemprop=\"subjectOf\" itemscope itemtype=\"https://schema.org/VideoObject\"\u003E\n\n\t\u003C!-- Card --\u003E\n\t\u003Cdiv class=\"card h-100 material-card has-spectrum-border-top__hover\" data-mount=\"click-group\"\u003E\n\t\t\u003C!-- Image --\u003E\n\t\t\t\t\t\u003Cdiv class=\"position-relative\"\u003E\n\t\t\t\t\u003Cimg width=\"390\" height=\"228\" src=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2025/12/ResearchForum_S2E2_Chenglong-Wang_1280x720-480x280.jpg\" class=\"card-img wp-post-image\" alt=\"Microsoft Research Forum | Chenglong Wang | Data Formulator: Vibe with your data, in control\" /\u003E\t\t\t\t\u003C!-- Duration --\u003E\n\t\t\t\t\u003Cspan class=\"duration--overlay badge font-weight-normal position-absolute bg-black text-white px-2 right-2 bottom-2 text-decoration-none\"\u003E\n\t\t\t\t\t\u003Cspan class=\"glyph-prepend glyph-prepend-video position-relative\" style=\"top: 2px\"\u003E\u003C/span\u003E\n\n\t\t\t\t\t\t\t\t\t\t\t\u003Cspan class=\"sr-only\"\u003EDuration\u003C/span\u003E\n\t\t\t\t\t\t \n\t\t\t\t\t\t06:40\t\t\t\t\t\t\t\t\t\u003C/span\u003E\n\t\t\t\u003C/div\u003E\n\t\t\t\t\t\t\t\u003Cmeta itemprop=\"thumbnail\" content=\"https://www.microsoft.com/en-us/research/wp-content/uploads/2025/12/ResearchForum_S2E2_Chenglong-Wang_1280x720-480x280.jpg\"\u003E\n\t\t\t\t\t\n\t\t\u003C!-- Card header --\u003E\n\t\t\u003Cdiv class=\"card-header mt-4 px-4\"\u003E\n\t\t\t\u003Ch3 itemprop=\"name\" class=\"mb-0 h4\"\u003E\n\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/video/data-formulator-vibe-with-your-data-in-control/\" data-bi-cN=\"Data Formulator: Vibe with your data, in control\" class=\"js-card-link icon-link icon-link--card-title text-body\" itemprop=\"url\" data-bi-type=\"video\" data-bi-tN=\"related-content-tabs\" aria-label=\"Play video entitled Data Formulator: Vibe with your data, in control\" icon_class=\"c-heading__icon\"\u003E\u003Cspan\u003EData Formulator: Vibe with your data, in control\u003C/span\u003E \u003Cspan class=\"glyph-in-link glyph-append glyph-append-chevron-right\" aria-hidden=\"true\"\u003E\u003C/span\u003E\u003C/a\u003E\t\t\t\u003C/h3\u003E\n\t\t\u003C/div\u003E\n\n\t\t\u003C!-- Card body --\u003E\n\t\t\u003Cdiv class=\"card-body p-4\"\u003E\n\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Video date\"\u003E\n\t\t\t\tDecember 9, 2025\t\t\t\u003C/p\u003E\n\n\t\t\t\t\t\t\t\u003Cdiv aria-label=\"Video speakers\"\u003E\n\t\t\t\t\t\n\t\t\t\t\t\t\u003Cp class=\"mb-1\"\u003E\n\t\t\t\t\t\t\t\u003Ca href=\"https://www.microsoft.com/en-us/research/people/keaster/\"\u003EKaren Easterbrook\u003C/a\u003E, \u003Ca href=\"https://www.microsoft.com/en-us/research/people/chenwang/\"\u003EChenglong Wang\u003C/a\u003E\t\t\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\t\t\t\t\u003C/div\u003E\n\t\t\t\n\t\t\t\t\t\t\t\u003Cp class=\"mb-1\" aria-label=\"Related event\"\u003E\n\t\t\t\t\tResearch Forum |\n\t\t\t\t\tSeason 2, Episode 2\t\t\t\t\u003C/p\u003E\n\t\t\t\t\t\u003C/div\u003E\n\t\u003C/div\u003E\n\u003C/div\u003E\n"}],"found_posts":56,"posts_per_page":12,"max_num_pages":5,"page":1,"current_selections":[{"id":"sort-by","name":"sort_by","count":0,"value":"most-recent","parent":{"id":"sort-by","label":"Sort by"},"slug":"sort_by","type":"sort-by","label":"Most recent","children":[],"selected":true,"default":true}],"facets":[{"id":"sort-by","label":"Sort by","slug":"sort-by","kind":"multiitemlist","items":[{"id":"sort-by","name":"sort_by","count":0,"value":"most-relevant","parent":{"id":"sort-by","label":"Sort by"},"slug":"sort_by","type":"sort-by","label":"Most relevant","children":[],"selected":false,"default":false},{"id":"sort-by","name":"sort_by","count":0,"value":"most-recent","parent":{"id":"sort-by","label":"Sort by"},"slug":"sort_by","type":"sort-by","label":"Most recent","children":[],"selected":true,"default":true},{"id":"sort-by","name":"sort_by","count":0,"value":"a-to-z","parent":{"id":"sort-by","label":"Sort by"},"slug":"sort_by","type":"sort-by","label":"Alphabetical: a to z","children":[],"selected":false,"default":false},{"id":"sort-by","name":"sort_by","count":0,"value":"z-to-a","parent":{"id":"sort-by","label":"Sort by"},"slug":"sort_by","type":"sort-by","label":"Alphabetical: z to a","children":[],"selected":false,"default":false}]},{"id":"msr-episode","label":"Episodes","slug":"msr-episode","kind":"multiitemlist","items":[{"id":"tax-269924","name":"facet[tax][msr-episode]","count":8,"value":"269924","parent":{"id":"msr-episode","label":"Episodes"},"slug":"msr-episode","type":"tax","label":"Season 1, Episode 1","children":[],"default":false,"selected":false},{"id":"tax-269925","name":"facet[tax][msr-episode]","count":7,"value":"269925","parent":{"id":"msr-episode","label":"Episodes"},"slug":"msr-episode","type":"tax","label":"Season 1, Episode 2","children":[],"default":false,"selected":false},{"id":"tax-269927","name":"facet[tax][msr-episode]","count":7,"value":"269927","parent":{"id":"msr-episode","label":"Episodes"},"slug":"msr-episode","type":"tax","label":"Season 1, Episode 4","children":[],"default":false,"selected":false},{"id":"tax-269928","name":"facet[tax][msr-episode]","count":7,"value":"269928","parent":{"id":"msr-episode","label":"Episodes"},"slug":"msr-episode","type":"tax","label":"Season 1, Episode 5","children":[],"default":false,"selected":false},{"id":"tax-269926","name":"facet[tax][msr-episode]","count":6,"value":"269926","parent":{"id":"msr-episode","label":"Episodes"},"slug":"msr-episode","type":"tax","label":"Season 1, Episode 3","children":[],"default":false,"selected":false},{"id":"tax-270235","name":"facet[tax][msr-episode]","count":6,"value":"270235","parent":{"id":"msr-episode","label":"Episodes"},"slug":"msr-episode","type":"tax","label":"Season 2, Episode 3","children":[],"default":false,"selected":false},{"id":"tax-270092","name":"facet[tax][msr-episode]","count":5,"value":"270092","parent":{"id":"msr-episode","label":"Episodes"},"slug":"msr-episode","type":"tax","label":"Season 2, Episode 1","children":[],"default":false,"selected":false},{"id":"tax-270184","name":"facet[tax][msr-episode]","count":5,"value":"270184","parent":{"id":"msr-episode","label":"Episodes"},"slug":"msr-episode","type":"tax","label":"Season 2, Episode 2","children":[],"default":false,"selected":false},{"id":"tax-270329","name":"facet[tax][msr-episode]","count":5,"value":"270329","parent":{"id":"msr-episode","label":"Episodes"},"slug":"msr-episode","type":"tax","label":"Season 2, Episode 4","children":[],"default":false,"selected":false}]},{"id":"msr-research-theme","label":"Research Themes","slug":"msr-research-theme","kind":"multiitemlist","items":[{"id":"tax-270111","name":"facet[tax][msr-research-theme]","count":18,"value":"270111","parent":{"id":"msr-research-theme","label":"Research Themes"},"slug":"msr-research-theme","type":"tax","label":"AI for all","children":[],"default":false,"selected":false},{"id":"tax-270109","name":"facet[tax][msr-research-theme]","count":11,"value":"270109","parent":{"id":"msr-research-theme","label":"Research Themes"},"slug":"msr-research-theme","type":"tax","label":"Mission on AI","children":[],"default":false,"selected":false},{"id":"tax-270110","name":"facet[tax][msr-research-theme]","count":11,"value":"270110","parent":{"id":"msr-research-theme","label":"Research Themes"},"slug":"msr-research-theme","type":"tax","label":"New advancement out of the lab","children":[],"default":false,"selected":false},{"id":"tax-270112","name":"facet[tax][msr-research-theme]","count":9,"value":"270112","parent":{"id":"msr-research-theme","label":"Research Themes"},"slug":"msr-research-theme","type":"tax","label":"Open research exploration","children":[],"default":false,"selected":false},{"id":"tax-270108","name":"facet[tax][msr-research-theme]","count":7,"value":"270108","parent":{"id":"msr-research-theme","label":"Research Themes"},"slug":"msr-research-theme","type":"tax","label":"Impact at scale","children":[],"default":false,"selected":false}]}],"filter_queries":[]},"classes":"btn btn-primary faceted-search__load-more","i18n":{"loadingStart":"Loading additional posts. Please wait.","loadingEnd":"Loading additional posts complete.","duetDatePicker":{"buttonLabel":"Choose date","placeholder":"YYYY-MM-DD","selectedDateMessage":"Selected date is","prevMonthLabel":"Previous month","nextMonthLabel":"Next month","monthSelectLabel":"Month","yearSelectLabel":"Year","closeLabel":"Close window","calendarHeading":"Choose a date","dayNames":["Sunday","Monday","Tuesday","Wednesday","Thursday","Friday","Saturday"],"monthNames":["January","February","March","April","May","June","July","August","September","October","November","December"],"monthNamesShort":["Jan","Feb","Mar","Apr","May","Jun","Jul","Aug","Sep","Oct","Nov","Dec"]},"validation":{"patternMismatch":"Please enter a date in YYYY-MM-DD format."}},"slot":"0","isSearch":"","defaultFacets":{"facet":[],"sort_by":"most-recent"},"theme_assets":"https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/","postId":"1140163","postTab":""};
//# sourceURL=faceted-search-js-extra
</script>
<script id="faceted-search-js" src="https://www.microsoft.com/en-us/research/wp-content/themes/microsoft-research-theme/assets/js/search.min.js?ver=e4d4a04b427ce70cdab880a927f749fbcbb9a332"></script>
<script id="wp-emoji-settings" type="application/json">
{"baseUrl":"https://s.w.org/images/core/emoji/17.0.2/72x72/","ext":".png","svgUrl":"https://s.w.org/images/core/emoji/17.0.2/svg/","svgExt":".svg","source":{"concatemoji":"https://www.microsoft.com/en-us/research/wp-includes/js/wp-emoji-release.min.js?ver=7.0.4"}}
</script>
<script type="module">
/*! This file is auto-generated */
var e="script#wp-emoji-settings",t=document.querySelector(e);if(!(t instanceof HTMLScriptElement))throw new Error("Element missing: "+e);const r=JSON.parse(t.text),s=(window._wpemojiSettings=r,"wpEmojiSettingsSupports"),o=["flag","emoji"];function i(e){try{var t={supportTests:e,timestamp:(new Date).valueOf()};sessionStorage.setItem(s,JSON.stringify(t))}catch(e){}}function c(e,t,n){e.clearRect(0,0,e.canvas.width,e.canvas.height),e.fillText(t,0,0);t=new Uint32Array(e.getImageData(0,0,e.canvas.width,e.canvas.height).data);e.clearRect(0,0,e.canvas.width,e.canvas.height),e.fillText(n,0,0);const r=new Uint32Array(e.getImageData(0,0,e.canvas.width,e.canvas.height).data);return t.every((e,t)=>e===r[t])}function p(e,t){e.clearRect(0,0,e.canvas.width,e.canvas.height),e.fillText(t,0,0);var n=e.getImageData(16,16,1,1);for(let e=0;e<n.data.length;e++)if(0!==n.data[e])return!1;return!0}function u(e,t,n,r){switch(t){case"flag":return n(e,"\ud83c\udff3\ufe0f\u200d\u26a7\ufe0f","\ud83c\udff3\ufe0f\u200b\u26a7\ufe0f")?!1:!n(e,"\ud83c\udde8\ud83c\uddf6","\ud83c\udde8\u200b\ud83c\uddf6")&&!n(e,"\ud83c\udff4\udb40\udc67\udb40\udc62\udb40\udc65\udb40\udc6e\udb40\udc67\udb40\udc7f","\ud83c\udff4\u200b\udb40\udc67\u200b\udb40\udc62\u200b\udb40\udc65\u200b\udb40\udc6e\u200b\udb40\udc67\u200b\udb40\udc7f");case"emoji":return!r(e,"\ud83e\u1fac8")}return!1}function f(e,t,n,r){let a;const s=(a="undefined"!=typeof WorkerGlobalScope&&self instanceof WorkerGlobalScope?new OffscreenCanvas(300,150):document.createElement("canvas")).getContext("2d",{willReadFrequently:!0}),o=(s.textBaseline="top",s.font="600 32px Arial",{});return e.forEach(e=>{o[e]=t(s,e,n,r)}),o}function a(e){var t=document.createElement("script");t.src=e,t.defer=!0,document.head.appendChild(t)}r.supports={everything:!0,everythingExceptFlag:!0},new Promise(t=>{let n=function(){try{var e=JSON.parse(sessionStorage.getItem(s));if("object"==typeof e&&"number"==typeof e.timestamp&&(new Date).valueOf()<e.timestamp+604800&&"object"==typeof e.supportTests)return e.supportTests}catch(e){}return null}();if(!n){if("undefined"!=typeof Worker&&"undefined"!=typeof OffscreenCanvas&&"undefined"!=typeof URL&&URL.createObjectURL&&"undefined"!=typeof Blob)try{var e="postMessage("+f.toString()+"("+[JSON.stringify(o),u.toString(),c.toString(),p.toString()].join(",")+"));",r=new Blob([e],{type:"text/javascript"});const a=new Worker(URL.createObjectURL(r),{name:"wpTestEmojiSupports"});return void(a.onmessage=e=>{i(n=e.data),a.terminate(),t(n)})}catch(e){}i(n=f(o,u,c,p))}t(n)}).then(e=>{for(const n in e)r.supports[n]=e[n],r.supports.everything=r.supports.everything&&r.supports[n],"flag"!==n&&(r.supports.everythingExceptFlag=r.supports.everythingExceptFlag&&r.supports[n]);var t;r.supports.everythingExceptFlag=r.supports.everythingExceptFlag&&!r.supports.flag,r.supports.everything||((t=r.source||{}).concatemoji?a(t.concatemoji):t.wpemoji&&t.twemoji&&(a(t.twemoji),a(t.wpemoji)))});
//# sourceURL=https://www.microsoft.com/en-us/research/wp-includes/js/wp-emoji-loader.min.js
</script>
</body>
</html>
Sitemap
Кол-во: 0
XML-карта сайта для поисковиков
?
Sitemap.xml помогает поисковику быстрее находить и индексировать страницы. Особенно важен для крупных сайтов и новых страниц, на которые ещё нет входящих ссылок.
Robots.txt не содержит ссылку на карту сайта. Рекомендуется добавить карту сайта и указать ссылку на нее в robots.txt.
Внутренние ссылки
Кол-во: 52
Ссылки на другие страницы своего сайта
?
Внутренние ссылки распределяют ссылочный вес между страницами и помогают поисковику обходить сайт. Пустые анкоры и ссылки на запрещённые robots.txt страницы — типичные ошибки.
Внутренних ссылок на странице 52 оптимально.
Внутренние ссылки не запрещены к индексации в robots.txt.
Показать внутренние ссылки
| Url | Анкор | Состояние | Анализировать |
|---|---|---|---|
| / |
Skip to main content
|
|
Анализировать url |
| /en-us/research/ |
<span>Research</span>
|
|
Анализировать url |
| /en-us/research/publications/ |
Publications
|
|
Анализировать url |
| /en-us/research/tools/ |
Code & data
|
|
Анализировать url |
| /en-us/research/people/ |
People
|
|
Анализировать url |
| /en-us/research/blog/ |
Microsoft Research blog
|
|
Анализировать url |
| /en-us/research/focus-area/ai-and-microsoft-research/ |
Artificial intelligence
|
|
Анализировать url |
| /en-us/research/research-area/audio-acoustics/ |
Audio & acoustics
|
|
Анализировать url |
| /en-us/research/research-area/computer-vision/ |
Computer vision
|
|
Анализировать url |
| /en-us/research/research-area/graphics-and-multimedia/ |
Graphics & multimedia
|
|
Анализировать url |
| /en-us/research/research-area/human-computer-interaction/ |
Human-computer interaction
|
|
Анализировать url |
| /en-us/research/research-area/human-language-technologies/ |
Human language technologies
|
|
Анализировать url |
| /en-us/research/research-area/search-information-retrieval/ |
Search & information retrieval
|
|
Анализировать url |
| /en-us/research/research-area/data-platform-analytics/ |
Data platforms and analytics
|
|
Анализировать url |
| /en-us/research/research-area/hardware-devices/ |
Hardware & devices
|
|
Анализировать url |
| /en-us/research/research-area/programming-languages-software-engineering/ |
Programming languages & software engineering
|
|
Анализировать url |
| /en-us/research/research-area/quantum/ |
Quantum computing
|
|
Анализировать url |
| /en-us/research/research-area/security-privacy-cryptography/ |
Security, privacy & cryptography
|
|
Анализировать url |
| /en-us/research/research-area/systems-and-networking/ |
Systems & networking
|
|
Анализировать url |
| /en-us/research/research-area/algorithms/ |
Algorithms
|
|
Анализировать url |
| /en-us/research/research-area/computational-sciences-mathematics/ |
Mathematics
|
|
Анализировать url |
| /en-us/research/research-area/ecology-environment/ |
Ecology & environment
|
|
Анализировать url |
| /en-us/research/research-area/economics/ |
Economics
|
|
Анализировать url |
| /en-us/research/research-area/medical-health-genomics/ |
Medical, health & genomics
|
|
Анализировать url |
| /en-us/research/research-area/social-sciences/ |
Social sciences
|
|
Анализировать url |
| /en-us/research/research-area/technology-for-emerging-markets/ |
Technology for emerging markets
|
|
Анализировать url |
| /en-us/research/academic-programs/ |
Academic programs
|
|
Анализировать url |
| /en-us/research/events-conferences/ |
Events & academic conferences
|
|
Анализировать url |
|
Microsoft Research Forum
|
|
Анализировать url | |
| /en-us/research/blog |
Microsoft Research blog
|
|
Анализировать url |
|
Microsoft Research Forum
|
|
Анализировать url | |
| /en-us/research/podcast/ |
Microsoft Research podcast
|
|
Анализировать url |
| /en-us/research/about-microsoft-research/ |
About Microsoft Research
|
|
Анализировать url |
| /en-us/research/careers/ |
Careers & internships
|
|
Анализировать url |
| /en-us/research/people/ |
People
|
|
Анализировать url |
| /en-us/research/microsoft-research-emeritus-program/ |
Emeritus program
|
|
Анализировать url |
| /en-us/research/news-and-awards/ |
News & awards
|
|
Анализировать url |
| /en-us/research/lab/microsoft-research-lab-africa-nairobi/ |
Africa
|
|
Анализировать url |
| /en-us/research/lab/microsoft-research-ai-for-science/ |
AI for Science
|
|
Анализировать url |
| /en-us/research/lab/ai-frontiers/ |
AI Frontiers
|
|
Анализировать url |
| /en-us/research/lab/microsoft-research-asia/ |
Asia-Pacific
|
|
Анализировать url |
| /en-us/research/lab/microsoft-research-cambridge/ |
Cambridge
|
|
Анализировать url |
| /en-us/research/lab/microsoft-health-futures/ |
Health Futures
|
|
Анализировать url |
| /en-us/research/lab/microsoft-research-india/ |
India
|
|
Анализировать url |
| /en-us/research/lab/microsoft-research-montreal/ |
Montreal
|
|
Анализировать url |
| /en-us/research/lab/microsoft-research-new-england/ |
New England
|
|
Анализировать url |
| /en-us/research/lab/microsoft-research-new-york/ |
New York City
|
|
Анализировать url |
| /en-us/research/lab/microsoft-research-redmond/ |
Redmond
|
|
Анализировать url |
| /en-us/research/lab/applied-sciences-group/ |
Applied Sciences
|
|
Анализировать url |
| /en-us/research/lab/mixed-reality-ai-lab-cambridge/ |
Mixed Reality & AI - Cambridge
|
|
Анализировать url |
| /en-us/research/lab/mixed-reality-ai-zurich/ |
Mixed Reality & AI - Zurich
|
|
Анализировать url |
|
Register: Research Forum
|
|
Анализировать url |
Внешние ссылки
Кол-во: 118
Ссылки на сторонние сайты
?
Исходящие внешние ссылки передают часть ссылочного веса на чужие сайты. Ссылки на авторитетные ресурсы безопасны; ссылки на мусорные сайты могут навредить репутации страницы.
Внешних ссылок на странице 118 слишком много. Спрячьте лишние ссылки в тег noindex или атрибут rel='nofollow'!
Показать первые 100 внешних ссылок
| Url | Анкор | Анализировать |
|---|---|---|
| microsoft.com |
Behind the Tech podcast
|
Анализировать url |
| info.microsoft.com |
Microsoft Research newsletter
|
Анализировать url |
| microsoft.com |
Microsoft Security
|
Анализировать url |
| azure.microsoft.com |
Azure
|
Анализировать url |
| dynamics.microsoft.com |
Dynamics 365
|
Анализировать url |
| microsoft.com |
Microsoft 365
|
Анализировать url |
| microsoft.com |
Microsoft Teams
|
Анализировать url |
| microsoft.com |
Windows 365
|
Анализировать url |
| microsoft.com |
Microsoft AI
|
Анализировать url |
| azure.microsoft.com |
Azure Space
|
Анализировать url |
| microsoft.com |
Mixed reality
|
Анализировать url |
| microsoft.com |
Microsoft HoloLens
|
Анализировать url |
| microsoft.com |
Microsoft Viva
|
Анализировать url |
| azure.microsoft.com |
Quantum computing
|
Анализировать url |
| microsoft.com |
Sustainability
|
Анализировать url |
| microsoft.com |
Education
|
Анализировать url |
| microsoft.com |
Automotive
|
Анализировать url |
| microsoft.com |
Financial services
|
Анализировать url |
| microsoft.com |
Government
|
Анализировать url |
| microsoft.com |
Healthcare
|
Анализировать url |
| microsoft.com |
Manufacturing
|
Анализировать url |
| microsoft.com |
Retail
|
Анализировать url |
| partner.microsoft.com |
Find a partner
|
Анализировать url |
| partner.microsoft.com |
Become a partner
|
Анализировать url |
| partner.microsoft.com |
Partner Network
|
Анализировать url |
| marketplace.microsoft.com |
Microsoft Marketplace
|
Анализировать url |
| microsoft.com |
Software companies
|
Анализировать url |
| blogs.microsoft.com |
Blog
|
Анализировать url |
| about.ads.microsoft.com |
Microsoft Advertising
|
Анализировать url |
| developer.microsoft.com |
Developer Center
|
Анализировать url |
| learn.microsoft.com |
Documentation
|
Анализировать url |
| microsoft.com |
Events
|
Анализировать url |
| microsoft.com |
Licensing
|
Анализировать url |
| learn.microsoft.com |
Microsoft Learn
|
Анализировать url |
| microsoft.com |
Microsoft Research
|
Анализировать url |
| microsoft.com |
View Sitemap
|
Анализировать url |
| aka.ms |
Register for the series
|
Анализировать url |
| microsoft.com |
Overview
|
Анализировать url |
| microsoft.com |
Past episodes
|
Анализировать url |
| microsoft.com |
<span>MagenticLite: A full-stack agentic experience powered by Small Models</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span>
|
Анализировать url |
| microsoft.com |
<span>Introducing GitHub Agentic Workflows: AI that runs your repo</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span>
|
Анализировать url |
| microsoft.com |
<span>Introducing Interwhen: Steering reasoning agents with real-time verification</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span>
|
Анализировать url |
| microsoft.com |
<span>New fine-tuning of language models: Match meaning, not tokens</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span>
|
Анализировать url |
| microsoft.com |
<span>Guiding the AI disruption to the Good Place</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span>
|
Анализировать url |
| aka.ms |
All Research Forum sessions
|
Анализировать url |
| microsoft.com |
<span>Microsoft at CHI 2026</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span>
|
Анализировать url |
| microsoft.com |
<span>Explore careers in research</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span>
|
Анализировать url |
| microsoft.com |
<span>Microsoft Research Blog</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span>
|
Анализировать url |
| microsoft.com |
<span>Microsoft at ICLR 2026</span> <span class="glyph-in-link glyph-append glyph-append-chevron-right" aria-hidden="true"></span>
|
Анализировать url |
| aka.ms |
Register for the series
|
Анализировать url |
| microsoft.com |
Event Code of Conduct
|
Анализировать url |
| x.com |
<span class="sr-only">Follow on X</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M18.42,14.009L27.891,3h-2.244l-8.224,9.559L10.855,3H3.28l9.932,14.455L3.28,29h2.244l8.684-10.095,6.936,10.095h7.576l-10.301-14.991h0Zm-3.074,3.573l-1.006-1.439L6.333,4.69h3.447l6.462,9.243,1.006,1.439,8.4,12.015h-3.447l-6.854-9.804h0Z"></path></svg>
|
Анализировать url |
| facebook.com |
<span class="sr-only">Like on Facebook</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M16,2c-7.732,0-14,6.268-14,14,0,6.566,4.52,12.075,10.618,13.588v-9.31h-2.887v-4.278h2.887v-1.843c0-4.765,2.156-6.974,6.835-6.974,.887,0,2.417,.174,3.043,.348v3.878c-.33-.035-.904-.052-1.617-.052-2.296,0-3.183,.87-3.183,3.13v1.513h4.573l-.786,4.278h-3.787v9.619c6.932-.837,12.304-6.74,12.304-13.897,0-7.732-6.268-14-14-14Z"></path></svg>
|
|
| linkedin.com |
<span class="sr-only">Follow on LinkedIn</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M26.111,3H5.889c-1.595,0-2.889,1.293-2.889,2.889V26.111c0,1.595,1.293,2.889,2.889,2.889H26.111c1.595,0,2.889-1.293,2.889-2.889V5.889c0-1.595-1.293-2.889-2.889-2.889ZM10.861,25.389h-3.877V12.87h3.877v12.519Zm-1.957-14.158c-1.267,0-2.293-1.034-2.293-2.31s1.026-2.31,2.293-2.31,2.292,1.034,2.292,2.31-1.026,2.31-2.292,2.31Zm16.485,14.158h-3.858v-6.571c0-1.802-.685-2.809-2.111-2.809-1.551,0-2.362,1.048-2.362,2.809v6.571h-3.718V12.87h3.718v1.686s1.118-2.069,3.775-2.069,4.556,1.621,4.556,4.975v7.926Z" fill-rule="evenodd"></path></svg>
|
Анализировать url |
| youtube.com |
<span class="sr-only">Subscribe on Youtube</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M31.331,8.248c-.368-1.386-1.452-2.477-2.829-2.848-2.496-.673-12.502-.673-12.502-.673,0,0-10.007,0-12.502,.673-1.377,.37-2.461,1.462-2.829,2.848-.669,2.512-.669,7.752-.669,7.752,0,0,0,5.241,.669,7.752,.368,1.386,1.452,2.477,2.829,2.847,2.496,.673,12.502,.673,12.502,.673,0,0,10.007,0,12.502-.673,1.377-.37,2.461-1.462,2.829-2.847,.669-2.512,.669-7.752,.669-7.752,0,0,0-5.24-.669-7.752ZM12.727,20.758V11.242l8.364,4.758-8.364,4.758Z" fill="currentColor"></path></svg>
|
Анализировать url |
| instagram.com |
<span class="sr-only">Follow on Instagram</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M10.202,2.098c-1.49,.07-2.507,.308-3.396,.657-.92,.359-1.7,.84-2.477,1.619-.776,.779-1.254,1.56-1.61,2.481-.345,.891-.578,1.909-.644,3.4-.066,1.49-.08,1.97-.073,5.771s.024,4.278,.096,5.772c.071,1.489,.308,2.506,.657,3.396,.359,.92,.84,1.7,1.619,2.477,.779,.776,1.559,1.253,2.483,1.61,.89,.344,1.909,.579,3.399,.644,1.49,.065,1.97,.08,5.771,.073,3.801-.007,4.279-.024,5.773-.095s2.505-.309,3.395-.657c.92-.36,1.701-.84,2.477-1.62s1.254-1.561,1.609-2.483c.345-.89,.579-1.909,.644-3.398,.065-1.494,.081-1.971,.073-5.773s-.024-4.278-.095-5.771-.308-2.507-.657-3.397c-.36-.92-.84-1.7-1.619-2.477s-1.561-1.254-2.483-1.609c-.891-.345-1.909-.58-3.399-.644s-1.97-.081-5.772-.074-4.278,.024-5.771,.096m.164,25.309c-1.365-.059-2.106-.286-2.6-.476-.654-.252-1.12-.557-1.612-1.044s-.795-.955-1.05-1.608c-.192-.494-.423-1.234-.487-2.599-.069-1.475-.084-1.918-.092-5.656s.006-4.18,.071-5.656c.058-1.364,.286-2.106,.476-2.6,.252-.655,.556-1.12,1.044-1.612s.955-.795,1.608-1.05c.493-.193,1.234-.422,2.598-.487,1.476-.07,1.919-.084,5.656-.092,3.737-.008,4.181,.006,5.658,.071,1.364,.059,2.106,.285,2.599,.476,.654,.252,1.12,.555,1.612,1.044s.795,.954,1.051,1.609c.193,.492,.422,1.232,.486,2.597,.07,1.476,.086,1.919,.093,5.656,.007,3.737-.006,4.181-.071,5.656-.06,1.365-.286,2.106-.476,2.601-.252,.654-.556,1.12-1.045,1.612s-.955,.795-1.608,1.05c-.493,.192-1.234,.422-2.597,.487-1.476,.069-1.919,.084-5.657,.092s-4.18-.007-5.656-.071M21.779,8.517c.002,.928,.755,1.679,1.683,1.677s1.679-.755,1.677-1.683c-.002-.928-.755-1.679-1.683-1.677,0,0,0,0,0,0-.928,.002-1.678,.755-1.677,1.683m-12.967,7.496c.008,3.97,3.232,7.182,7.202,7.174s7.183-3.232,7.176-7.202c-.008-3.97-3.233-7.183-7.203-7.175s-7.182,3.233-7.174,7.203m2.522-.005c-.005-2.577,2.08-4.671,4.658-4.676,2.577-.005,4.671,2.08,4.676,4.658,.005,2.577-2.08,4.671-4.658,4.676-2.577,.005-4.671-2.079-4.676-4.656h0"></path></svg>
|
|
| microsoft.com |
<span class="sr-only">Subscribe to our RSS feed</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><circle cx="6.566" cy="25.434" r="3.566" fill="currentColor"></circle><path d="M20.234,29h-5.051c0-6.728-5.454-12.183-12.183-12.183h0v-5.051c9.518,0,17.234,7.716,17.234,17.234Z" fill="currentColor"></path><path d="M23.8,29c0-11.488-9.312-20.8-20.8-20.8V3c14.359,0,26,11.641,26,26h-5.2Z" fill="currentColor"></path></svg>
|
Анализировать url |
| x.com |
<span class="sr-only">Share on X</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M18.42,14.009L27.891,3h-2.244l-8.224,9.559L10.855,3H3.28l9.932,14.455L3.28,29h2.244l8.684-10.095,6.936,10.095h7.576l-10.301-14.991h0Zm-3.074,3.573l-1.006-1.439L6.333,4.69h3.447l6.462,9.243,1.006,1.439,8.4,12.015h-3.447l-6.854-9.804h0Z"></path></svg>
|
Анализировать url |
| facebook.com |
<span class="sr-only">Share on Facebook</span>
<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" viewbox="0 0 32 32" aria-hidden="true" class="icon-social"><path d="M16,2c-7.732,0-14,6.268-14,14,0,6.566,4.52,12.075,10.618,13.588v-9.31h-2.887v-4.278h2.887v-1.843c0-4.765,2.156-6.974,6.835-6.974,.887,0,2.417,.174,3.043,.348v3.878c-.33-.035-.904-.052-1.617-.052-2.296,0-3.183,.87-3.183,3.13v1.513h4.573l-.786,4.278h-3.787v9.619c6.932-.837,12.304-6.74,12.304-13.897,0-7.732-6.268-14-14-14Z"></path></svg>
|
|
| microsoft.com |
Surface Pro
|
Анализировать url |
| microsoft.com |
Surface Laptop
|
Анализировать url |
| microsoft.com |
Surface Laptop Ultra
|
Анализировать url |
| microsoft.com |
Surface RTX Spark Dev Box
|
Анализировать url |
| microsoft.com |
Copilot for organizations
|
Анализировать url |
| microsoft.com |
Copilot for personal use
|
Анализировать url |
| microsoft.com |
Explore Microsoft products
|
Анализировать url |
| microsoft.com |
Windows 11 apps
|
Анализировать url |
| account.microsoft.com |
Account profile
|
Анализировать url |
| microsoft.com |
Download Center
|
Анализировать url |
| go.microsoft.com |
Microsoft Store support
|
Анализировать url |
| microsoft.com |
Returns
|
Анализировать url |
| microsoft.com |
Order tracking
|
Анализировать url |
| microsoft.com |
Certified Refurbished
|
Анализировать url |
| microsoft.com |
Microsoft Store Promise
|
Анализировать url |
| microsoft.com |
Flexible Payments
|
Анализировать url |
| microsoft.com |
Microsoft in education
|
Анализировать url |
| microsoft.com |
Devices for education
|
Анализировать url |
| microsoft.com |
Microsoft Teams for Education
|
Анализировать url |
| microsoft.com |
Microsoft 365 Education
|
Анализировать url |
| microsoft.com |
How to buy for your school
|
Анализировать url |
| education.microsoft.com |
Educator training and development
|
Анализировать url |
| microsoft.com |
Deals for students and parents
|
Анализировать url |
| microsoft.com |
AI for education
|
Анализировать url |
| microsoft.com |
Microsoft AI
|
Анализировать url |
| microsoft.com |
Microsoft Security
|
Анализировать url |
| microsoft.com |
Dynamics 365
|
Анализировать url |
| microsoft.com |
Microsoft 365
|
Анализировать url |
| microsoft.com |
Microsoft Power Platform
|
Анализировать url |
| microsoft.com |
Microsoft Teams
|
Анализировать url |
| microsoft.com |
Microsoft 365 Copilot
|
Анализировать url |
| microsoft.com |
Small Business
|
Анализировать url |
| azure.microsoft.com |
Azure
|
Анализировать url |
| developer.microsoft.com |
Microsoft Developer
|
Анализировать url |
| learn.microsoft.com |
Microsoft Learn
|
Анализировать url |
| microsoft.com |
Support for AI marketplace apps
|
Анализировать url |
| techcommunity.microsoft.com |
Microsoft Tech Community
|
Анализировать url |
| marketplace.microsoft.com |
Microsoft Marketplace
|
Анализировать url |
| microsoft.com |
Software companies
|
Анализировать url |
| visualstudio.microsoft.com |
Visual Studio
|
Анализировать url |
| careers.microsoft.com |
Careers
|
Анализировать url |
Конкуренты Готовность: 0%
Конкуренты в Яндексе
Кол-во: 0
Топ сайтов-конкурентов в Яндексе
?
Сайты, чаще всего появляющиеся в ТОПе Яндекса по запросам из семантического ядра этой страницы.
Мы не нашли у вас конкурентов в Яндексе. Сайт или очень молодой или плохо продвигается.
Конкурентов в ТОП-10 Яндекса не нашлось.
Конкуренты в Google
Кол-во: 0
Топ сайтов-конкурентов в Google
?
Сайты, чаще всего появляющиеся в ТОПе Google по запросам из семантического ядра этой страницы.
Конкуренты в Google тоже не найдены. Займитесь продвижением сайта!
Конкурентов в ТОП-10 Google не нашлось.
ЗоЗПП: права потребителей Готовность: 100%
Нарушения
Не выявлены
Признаков дистанционной продажи товаров (интернет-магазина) не обнаружено — требования ЗоЗПП о раскрытии информации продавца к сайту не применяются. Нарушений нет.
ФЗ-149: рекомендательные технологии Готовность: 100%
Нарушения
Не выявлены
Рекомендательные блоки («с этим покупают», «похожие товары» и т.п.) на сайте не обнаружены — требования ст. 10.7 ФЗ-149 к сайту не применяются. Нарушений нет.
ФЗ-38: реклама Готовность: 100%
Нарушения
Не выявлены
Рекламных тематик с обязательными оговорками (медицина, БАД, кредиты и займы, новостройки) на сайте не обнаружено. Нарушений нет.
ФЗ-436: защита детей Готовность: 100%
Нарушения
Не выявлены
Признаков информационной продукции (новости, видео, книги, игры, курсы) не обнаружено — обязательная возрастная маркировка по ФЗ-436 сайту не требуется. Нарушений нет.
Вердикт
Сайт researchforum.microsoft.com не совсем готов к продвижению (процент готовности лишь 31%). Для попадания в ТОПы поисковых систем нужно:
Постарайтесь исправить общие ошибки.
Исправьте ошибки в мета-тегах.
Исправьте ошибки индексации.
Поделитесь с друзьями: