Анализ сайта jackgethome.com
Основное Готовность: 60%
Домен
jackgethome.com
Состояние доменного имени
?
Проверяем корректность доменного имени и наличие технических проблем на уровне домена.
Домен второго уровня идеален для продвижения.
Отличный запоминающийся домен.
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 4 слишком большое! Проверьте настройки сайта и веб-сервера!
Цепочка редиректов:
http://jackgethome.com
301 MovedPermanently
https://jackgethome.com/
301 MovedPermanently
https://www.jackgethome.com/
301 MovedPermanently
https://biechi.github.io/
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 06.11.2026 22:44:25.
Включён HSTS (Strict-Transport-Security) — защита от подмены на http.
HTTP автоматически перенаправляется на HTTPS.
Поздравляем! Сайт не содержится в реестре РКН.
Кодировка
utf-8
Кодировка символов страницы
?
Стандарт — UTF-8. Неправильная кодировка вызывает нечитаемые символы и мешает поисковику корректно распознать текст страницы.
Указана кодировка на странице utf-8.
Язык
en
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.
Скорость загрузки
~1,50сек
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 1,50сек превышает 1 секунду. Желательно улучшить работу сайта!
Объем документа
53Кб
Размер HTML-кода страницы
?
Слишком большой HTML замедляет парсинг браузером и сканирование поисковым роботом. Рекомендуется не более 200 Кб.
Объем html-документа 53Кб оптимален.
Структура html-документа корректна.
Ресурсы
Ресурсы: 27
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 27 много для одной страницы. Приемлемо до 10. Проведите оптимизацию файлов ресурсов!
Показать полный список ресурсов
| Тип | Название | Значение |
|---|---|---|
| stylesheet | /assets/css/bootstrap.min.css?a26b1b0e2d9c9dd2e5ed17c987a37731 | |
| stylesheet | https://cdn.jsdelivr.net/npm/mdbootstrap@4.20.0/css/mdb.min.css | |
| stylesheet | /assets/css/academicons.min.css?9dbe0c76bacd463500b79ed6a86e27ac | |
| stylesheet | text/css | https://fonts.googleapis.com/css?family=Roboto:300,400,500,700|Roboto+Slab:100,300,400,500,700|Material+Icons&display=swap |
| stylesheet | /assets/css/jekyll-pygments-themes-github.css?591dab5a4e56573bf4ef7fd332894c99 | |
| stylesheet | /assets/css/main.css?4b1b5d9b313aa29ae3096edd38c419b2 | |
| stylesheet | /assets/css/jekyll-pygments-themes-native.css?5847e5ed4a4568527aa6cfab446049ca | |
| js | /assets/js/theme.js?bd888c560287cd675855c7662a167c4a | |
| js | https://cdn.jsdelivr.net/npm/jquery@3.6.0/dist/jquery.min.js | |
| js | /assets/js/bootstrap.bundle.min.js | |
| js | https://cdn.jsdelivr.net/npm/mdbootstrap@4.20.0/js/mdb.min.js | |
| js | https://cdn.jsdelivr.net/npm/masonry-layout@4.2.2/dist/masonry.pkgd.min.js | |
| js | https://cdn.jsdelivr.net/npm/imagesloaded@5.0.0/imagesloaded.pkgd.min.js | |
| js | text/javascript | /assets/js/masonry.js |
| js | https://cdn.jsdelivr.net/npm/medium-zoom@1.1.0/dist/medium-zoom.min.js | |
| js | /assets/js/zoom.js?3e7054dc4d3e3dd8f0731a48453e618e | |
| js | /assets/js/sidenote.js?524e5ded3e4d9d041a39f9b243e5e260 | |
| js | /assets/js/no_defer.js?2781658a0a2b13ed609542042a859126 | |
| js | /assets/js/common.js?0012839f98c37f202100293077bc9852 | |
| js | text/javascript | /assets/js/copy_code.js?3577194613afa04501eb52f8f4164de9 |
| js | /assets/js/jupyter_new_tab.js?d9f17b6adc2311cbabd747f4538bb15f | |
| js | https://d1bxh8uas1mnw7.cloudfront.net/assets/embed.js | |
| js | https://badge.dimensions.ai/badge.js | |
| js | text/javascript | https://cdn.jsdelivr.net/npm/mathjax@3.2.0/es5/tex-mml-chtml.min.js |
| js | https://cdnjs.cloudflare.com/polyfill/v3/polyfill.min.js?features=es6 | |
| js | /assets/js/vanilla-back-to-top.min.js?982c80efc910ea9a6203400dcbd0a3af | |
| js | /assets/js/shortcut-key.js |
Серверные заголовки
Кол-во: 18
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 18шт. Подробнее про серверные заголовки.
Показать полный список серверных заголовков
| Ключ | Значение |
|---|---|
| Server | GitHub.com |
| Access-Control-Allow-Origin | * |
| Strict-Transport-Security | max-age=31556952 |
| ETag | "6a862d18-d51c" |
| Cache-Control | max-age=600 |
| x-proxy-cache | MISS |
| x-github-request-id | 1C4A:3E1B:586E8E:59114B:6A88DD7F |
| x-github-edge-region | fra |
| Accept-Ranges | bytes |
| Age | 0 |
| Date | Sat, 22 Aug 2026 01:38:43 GMT |
| Via | 1.1 varnish |
| X-Served-By | cache-fra-eddf8230194-FRA |
| X-Cache | HIT |
| x-cache-hits | 0 |
| x-timer | S1787362724.675293,VS0,VE121 |
| Vary | Accept-Encoding |
| x-fastly-request-id | 64de0ee663ada0a8798a8cd28861abeb9183e2ae |
CMS
Не определена
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.
CMS не определена. Вероятно, сайт самописный либо движок надёжно скрыт. Это не ошибка.
Веб-сервер
GitHub.com
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.
В заголовке Server указано: GitHub.com.
Мета-теги Готовность: 19%
Title
Jack (Hao) Bai
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.
Необходимо увеличить число символов в title (текущее значение мало: 16, минимум: 25, оптимально: от 40 до 45)
Дублей словоформ в title не найдено.
Description
Homepage of Jack (Hao) Bai.
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.
Необходимо увеличить число символов в description (текущее значение мало: 28, минимум: 80, оптимально: от 120 до 130)
Keywords
Jack Bai, Hao Bai, Homepage
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.
Keywords установлены.
Канонический Url
https://jackgethome.com/
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.
Пропишите корректный канонический Url!
Robots
Ошибок нет
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 не указан. Страница свободна для индексации.
Адаптивность
width=device-width, initial-scale=1, shrink-to-fit=no
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
Не найдено
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 не задана. Страница не оптимизирована под социальные сети. Мета-теги с разметкой Og помогают социальным роботам лучше структурировать Ваш сайт.
Все мета-теги
Кол-во: 4
Полный список мета-тегов страницы
?
Таблица всех meta-тегов, включая нестандартные. Позволяет найти опечатки, дубли и лишние теги.
Найдены мета-теги 4шт. 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, shrink-to-fit=no |
| name | author | Jack (Hao) Bai |
| name | description | Homepage of Jack (Hao) Bai. |
| name | keywords | Jack Bai, Hao Bai, Homepage |
Оптимизация Готовность: 80%
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
Ошибок нет
Объём и качество текстового содержимого
?
Анализирует объём полезного текста на странице. Слишком мало — страница может считаться малополезной. Слишком много — ухудшается читаемость и восприятие.
Слова из title 3 встречаются в тексте достаточно.
Абзацев с текстом 15 достаточно.
Среднее число слов в абзаце 132 достаточно.
Кол-во знаков контента 20304 на странице оптимально.
Кол-во слов 2981 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> 5. Это хорошо.
Тошнота
5,48
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 5. Измените текст страницы!
Академич. тошнота
28,65%
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 |
|---|---|---|
| agents | 30 | 1,01% |
| training | 18 | 0,60% |
| learning | 17 | 0,57% |
| models | 15 | 0,50% |
| approach | 15 | 0,50% |
| policy | 14 | 0,47% |
| performance | 10 | 0,34% |
| improvement | 10 | 0,34% |
| language | 10 | 0,34% |
| td-learning | 10 | 0,34% |
| multi-step | 9 | 0,30% |
| research | 8 | 0,27% |
| reasoning | 8 | 0,27% |
| reinforcement | 8 | 0,27% |
| vision-language | 7 | 0,23% |
| visual | 7 | 0,23% |
| webgym | 7 | 0,23% |
| framework | 7 | 0,23% |
| scaling | 7 | 0,23% |
| optimization | 7 | 0,23% |
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 116 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 4 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://jackgethome.com/robots.txt
301 MovedPermanently
https://jackgethome.com/robots.txt
301 MovedPermanently
https://www.jackgethome.com/robots.txt
301 MovedPermanently
https://biechi.github.io/
200 OK
Показать содержимое robots.txt
<!DOCTYPE html> <html lang="en"> <head> <meta http-equiv="Content-Type" content="text/html; charset=UTF-8"> <meta charset="utf-8"> <meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no"> <meta http-equiv="X-UA-Compatible" content="IE=edge"> <title> Jack (Hao) Bai </title> <meta name="author" content="Jack (Hao) Bai"> <meta name="description" content="Homepage of Jack (Hao) Bai. "> <meta name="keywords" content="Jack Bai, Hao Bai, Homepage"> <link rel="stylesheet" href="/assets/css/bootstrap.min.css?a26b1b0e2d9c9dd2e5ed17c987a37731"> <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/mdbootstrap@4.20.0/css/mdb.min.css" integrity="sha256-jpjYvU3G3N6nrrBwXJoVEYI/0zw8htfFnhT9ljN3JJw=" crossorigin="anonymous"> <link defer rel="stylesheet" href="/assets/css/academicons.min.css?9dbe0c76bacd463500b79ed6a86e27ac"> <link defer rel="stylesheet" type="text/css" href="https://fonts.googleapis.com/css?family=Roboto:300,400,500,700|Roboto+Slab:100,300,400,500,700|Material+Icons&display=swap"> <link defer rel="stylesheet" href="/assets/css/jekyll-pygments-themes-github.css?591dab5a4e56573bf4ef7fd332894c99" media="" id="highlight_theme_light"> <link rel="shortcut icon" href="/assets/img/icon.png?1b13e4f317a879116f820d24b9a54d75"> <link rel="stylesheet" href="/assets/css/main.css?4b1b5d9b313aa29ae3096edd38c419b2"> <link rel="canonical" href="https://jackgethome.com/"> <script src="/assets/js/theme.js?bd888c560287cd675855c7662a167c4a"></script> <link defer rel="stylesheet" href="/assets/css/jekyll-pygments-themes-native.css?5847e5ed4a4568527aa6cfab446049ca" media="none" id="highlight_theme_dark"> <script>initTheme();</script> </head> <body class="fixed-top-nav sticky-bottom-footer"> <header> <nav id="navbar" class="navbar navbar-light navbar-expand-sm fixed-top" role="navigation"> <div class="container"> <button class="navbar-toggler collapsed ml-auto" type="button" data-toggle="collapse" data-target="#navbarNav" aria-controls="navbarNav" aria-expanded="false" aria-label="Toggle navigation"> <span class="sr-only">Toggle navigation</span> <span class="icon-bar top-bar"></span> <span class="icon-bar middle-bar"></span> <span class="icon-bar bottom-bar"></span> </button> <div class="collapse navbar-collapse text-right" id="navbarNav"> <ul class="navbar-nav ml-auto flex-nowrap"> <li class="nav-item active"> <a class="nav-link" href="/">about <span class="sr-only">(current)</span> </a> </li> <li class="toggle-container"> <button id="light-toggle" title="Change theme"> <i class="ti ti-sun-moon" id="light-toggle-system"></i> <i class="ti ti-moon-filled" id="light-toggle-dark"></i> <i class="ti ti-sun-filled" id="light-toggle-light"></i> </button> </li> </ul> </div> </div> </nav> <progress id="progress" value="0"> <div class="progress-container"> <span class="progress-bar"></span> </div> </progress> </header> <div class="container mt-5" role="main"> <div class="post"> <header class="post-header"> <h1 class="post-title" style="text-align: left;"> Jack (Hao) Bai </h1> <p class="desc" style="text-align: left;">haob2 AT illinois DOT edu</p> </header> <article> <div class="profile float-right"> <figure> <picture> <img src="/assets/img/professional.png?664e8c9d145d5e6328386476e04bdb3e" class="img-fluid z-depth-1 rounded" width="100%" height="auto" style=" " alt="professional.png" title="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> <div class="contact-icons" style="text-align: center; font-size: 36px;"> <style>.contact-icons a{color:#0272ce;text-decoration:none}.contact-icons a:hover{color:#b507ac}</style> <a href="https://scholar.google.com/citations?user=720Ix7QAAAAJ" title="Google Scholar" rel="external nofollow noopener" target="_blank"><i class="ai ai-google-scholar"></i></a> <a href="https://github.com/biechi" title="GitHub" rel="external nofollow noopener" target="_blank"><i class="fa-brands fa-github"></i></a> <a href="https://www.linkedin.com/in/jackgethome" title="LinkedIn" rel="external nofollow noopener" target="_blank"><i class="fa-brands fa-linkedin"></i></a> <a href="https://twitter.com/jackbot_cs" title="X" rel="external nofollow noopener" target="_blank"><i class="fa-brands fa-x-twitter"></i></a> <a href="https://www.zhihu.com/people/jackgethome" title="Zhihu" rel="external nofollow noopener" target="_blank"><i class="fa-brands fa-zhihu"></i></a> </div> </picture> </figure> </div> <div class="clearfix"> <p>Hi there! I’m Jack. I’m a second-year Ph.D. student at UIUC CS, advised by Prof. <a href="https://tongzhang-ml.org/" rel="external nofollow noopener" target="_blank">Tong Zhang</a>. I work closely with Prof. <a href="https://aviralkumar2907.github.io/" rel="external nofollow noopener" target="_blank">Aviral Kumar</a> @ CMU MLD. I am a part-time research intern at NVIDIA Research, managed by Jan Kautz.</p> <p>Recently, I research on fundamental questions on <span class="hover-note" data-note="I don't believe marginalizing vision out from reasoning is a long-term solution">vision-language</span> model <span class="hover-note" data-note="Under multi-step setting, reasoning has many different properties from those in single step settings">reasoning in multi-step environments</span>, modernly named “agents”, with <span class="hover-note" data-note='I am particularly interested in step-level value functions, especially Q functions, for sample efficiency ("short-cuts")'>reinforcement learning</span>. I tackle problems with both empirical insights and theoretical considerations.</p> <p>I was previously a MS student in Computer Science at UIUC, advised by Prof. <a href="https://nanjiang.cs.illinois.edu/" rel="external nofollow noopener" target="_blank">Nan Jiang</a>, during which time I had the fortune to visit UC Berkeley, advised by <a href="https://people.eecs.berkeley.edu/~svlevine/" rel="external nofollow noopener" target="_blank">Sergey Levine</a>, and a research intern at Microsoft Research. I received my dual undergrad degree from UIUC and Zhejiang University.</p> <p>In my free time, I study music theory, majoring in chord progression.</p> <p>A public up-to-date <strong>resume</strong> can be found <a href="/assets/pdf/resume.pdf">here</a>.</p> </div> <div class="text-center mt-4 mb-4"> <img src="/assets/img/logo.png" alt="Logo" class="img-fluid" style="max-width: 400px;"> </div> <h2> <a href="/news/" style="color: inherit">News</a> </h2> <div class="news"> <div class="table-responsive" style="max-height: 60vw"> <table class="table table-sm table-borderless"> <tr> <th scope="row" style="width: 20%">Jun 08, 2026</th> <td> Our new paper, <strong>AsyncWebRL: Efficient Multi-Step RL for Visual Web Agents</strong>, is now available on <a href="https://arxiv.org/abs/2606.05597" rel="external nofollow noopener" target="_blank">arXiv</a>. Beyond asynchronous training, we find that the per-trajectory length-normalization term in multi-step GRPO substantially changes RL dynamics: it weakens the negative gradient on long failed trajectories and encourages unnecessarily verbose behavior. Replacing it with a constant normalizer improves trajectory efficiency while preserving success. </td> </tr> <tr> <th scope="row" style="width: 20%">Mar 08, 2026</th> <td> Our paper <strong>WebGym</strong> has been accepted to <strong>CVPR 2026</strong>! Check out the paper on <a href="https://arxiv.org/abs/2601.02439" rel="external nofollow noopener" target="_blank">ArXiv</a> and the <a href="https://webgym-website.github.io/webgym.github.io/" rel="external nofollow noopener" target="_blank">project page</a>. </td> </tr> <tr> <th scope="row" style="width: 20%">Jan 09, 2026</th> <td> Today, we proudly announce the release of <strong>WebGym</strong>, the largest yet open-source RL training environment for visual web agents. The preprint can be accessed at <a href="https://arxiv.org/abs/2601.02439" rel="external nofollow noopener" target="_blank">ArXiv</a>. We proposed (1) the RL framework with highest rollout speed, (2) recipe that supports training agents on long-horizon tasks, and (3) scaling dimensions that effectively improves the RL performance with the task set proposed. </td> </tr> </table> </div> </div> <h2>Research Blogs</h2> <div class="research-timeline" style="margin: 0.5rem 0 1.5rem; position: relative;"> <style>.timeline-bar{position:relative;height:3px;background:var(--global-divider-color,#e8e8e8);border-radius:2px;margin:2.5rem 1rem .5rem}.timeline-dot{position:absolute;top:50%;transform:translate(-50%,-50%);width:10px;height:10px;border-radius:50%;background:var(--global-theme-color);cursor:pointer;transition:transform .2s ease,box-shadow .2s ease;z-index:2}.timeline-dot:hover{transform:translate(-50%,-50%) scale(1.5);box-shadow:0 0 0 4px color-mix(in srgb,var(--global-theme-color) 20%,transparent)}.timeline-dot .timeline-tooltip{position:absolute;bottom:calc(100% + 8px);left:50%;transform:translateX(-50%);white-space:nowrap;font-size:.72rem;color:var(--global-text-color);opacity:0;pointer-events:none;transition:opacity .2s ease;background:var(--global-bg-color);padding:3px 8px;border-radius:4px;box-shadow:0 1px 4px rgba(0,0,0,0.12);line-height:1.4;text-align:center}.timeline-dot:hover .timeline-tooltip{opacity:1}.timeline-year-label{position:absolute;top:10px;transform:translateX(-50%);font-size:.7rem;color:var(--global-text-color-light,#999);letter-spacing:.03em}.timeline-year-tick{position:absolute;top:50%;transform:translate(-50%,-50%);width:1px;height:10px;background:var(--global-divider-color,#e8e8e8);z-index:1}.timeline-milestone{position:absolute;top:50%;transform:translate(-50%,-50%);width:8px;height:8px;border-radius:50%;background:#e74c3c;cursor:default;transition:transform .2s ease,box-shadow .2s ease;z-index:2;border:1.5px solid rgba(255,255,255,0.8)}.timeline-milestone:hover{transform:translate(-50%,-50%) scale(1.5);box-shadow:0 0 0 4px rgba(231,76,60,0.2)}.timeline-milestone .timeline-tooltip{position:absolute;bottom:calc(100% + 8px);left:50%;transform:translateX(-50%);white-space:nowrap;font-size:.72rem;color:var(--global-text-color);opacity:0;pointer-events:none;transition:opacity .2s ease;background:var(--global-bg-color);padding:3px 8px;border-radius:4px;box-shadow:0 1px 4px rgba(0,0,0,0.12);line-height:1.4;text-align:center}.timeline-milestone:hover .timeline-tooltip{opacity:1}</style> <div class="timeline-bar"> <div class="timeline-year-tick" style="left: 2%;"></div> <span class="timeline-year-label" style="left: 2%;">2023</span> <div class="timeline-year-tick" style="left: 25%;"></div> <span class="timeline-year-label" style="left: 25%;">2024</span> <div class="timeline-year-tick" style="left: 49%;"></div> <span class="timeline-year-label" style="left: 49%;">2025</span> <div class="timeline-year-tick" style="left: 73%;"></div> <span class="timeline-year-label" style="left: 73%;">2026</span> <div class="timeline-year-tick" style="left: 97%;"></div> <span class="timeline-year-label" style="left: 97%;">2027</span> <div class="timeline-milestone" style="left: 0%;"> <span class="timeline-tooltip">Nov 2022<br>ChatGPT released</span> </div> <div class="timeline-milestone" style="left: 50%;"> <span class="timeline-tooltip">Jan 2025<br>DeepSeek-R1 released</span> </div> <div class="timeline-milestone" style="left: 66%;"> <span class="timeline-tooltip">Sep 2025<br>Qwen3-VL released</span> </div> <a href="/blog/b4c5k6/" class="timeline-dot" style="left: 9%;"> <span class="timeline-tooltip">Apr 2023<br>Backpropagation</span> </a> <a href="/blog/3b5b97/" class="timeline-dot" style="left: 12%;"> <span class="timeline-tooltip">Jun 2023<br>Self-Attention Layer and The Transformers Architecture</span> </a> <a href="/blog/9c4e2f/" class="timeline-dot" style="left: 18%;"> <span class="timeline-tooltip">Sep 2023<br>RoPE and M-RoPE: Rotation, Decay, and Multimodal Axes</span> </a> <a href="/blog/m1o2e3/" class="timeline-dot" style="left: 24%;"> <span class="timeline-tooltip">Dec 2023<br>Mixture of Experts Explained</span> </a> <a href="/blog/be11m4n/" class="timeline-dot" style="left: 28%;"> <span class="timeline-tooltip">Feb 2024<br>Bellman Operator Identities</span> </a> <a href="/blog/82d5b0/" class="timeline-dot" style="left: 30%;"> <span class="timeline-tooltip">Mar 2024<br>The Policy Gradient Family: PG, PPO, and AC</span> </a> <a href="/blog/p1i2t3/" class="timeline-dot" style="left: 32%;"> <span class="timeline-tooltip">Apr 2024<br>Policy Improvement Theorem</span> </a> <a href="/blog/035eca/" class="timeline-dot" style="left: 35%;"> <span class="timeline-tooltip">May 2024<br>Importance Sampling: Why and How</span> </a> <a href="/blog/9e7104/" class="timeline-dot" style="left: 36%;"> <span class="timeline-tooltip">Jun 2024<br>LLM Optimization Basics: Time</span> </a> <a href="/blog/e31c7f/" class="timeline-dot" style="left: 39%;"> <span class="timeline-tooltip">Aug 2024<br>LLM Optimization Basics: Memory</span> </a> <a href="/blog/89b69e/" class="timeline-dot" style="left: 44%;"> <span class="timeline-tooltip">Oct 2024<br>RL on Language under Single-step Settings</span> </a> <a href="/blog/d01fde/" class="timeline-dot" style="left: 54%;"> <span class="timeline-tooltip">Mar 2025<br>Can Language Models Be Critic Functions?</span> </a> <a href="/blog/76c93a/" class="timeline-dot" style="left: 59%;"> <span class="timeline-tooltip">May 2025<br>Policy Optimization without a Critic: The GRPO Family</span> </a> <a href="/blog/994498/" class="timeline-dot" style="left: 62%;"> <span class="timeline-tooltip">Jul 2025<br>Are Multi-step Agents Overthinking?</span> </a> <a href="/blog/a897ea/" class="timeline-dot" style="left: 63%;"> <span class="timeline-tooltip">Aug 2025<br>Challenges in Scaling Q-Learning</span> </a> <a href="/blog/19b81f/" class="timeline-dot" style="left: 65%;"> <span class="timeline-tooltip">Sep 2025<br>Pretraining, Post-training, and Test-Time Reasoning</span> </a> <a href="/blog/823eeb/" class="timeline-dot" style="left: 67%;"> <span class="timeline-tooltip">Oct 2025<br>Position: Why Web is a Good Environment to Study RL?</span> </a> <a href="/blog/wx-defense/" class="timeline-dot" style="left: 70%;"> <span class="timeline-tooltip">Nov 2025<br>Adaptive Sampling and Curriculum Methods</span> </a> <a href="/blog/e2b7d4/" class="timeline-dot" style="left: 72%;"> <span class="timeline-tooltip">Dec 2025<br>Autoregressive Embedding Models: Training, Attention, and Performance</span> </a> <a href="/blog/a7c3e1/" class="timeline-dot" style="left: 73%;"> <span class="timeline-tooltip">Jan 2026<br>How to Use Privileged Information in RL: On-policy Distillation</span> </a> <a href="/blog/690518/" class="timeline-dot" style="left: 76%;"> <span class="timeline-tooltip">Feb 2026<br>Generalizable Value Functions and Introverted Intuition (Ni)</span> </a> <a href="/blog/f1m2r3/" class="timeline-dot" style="left: 77%;"> <span class="timeline-tooltip">Mar 2026<br>What Does Flow-Matching Bring to Deep RL?</span> </a> <a href="/blog/w4m407/" class="timeline-dot" style="left: 79%;"> <span class="timeline-tooltip">Apr 2026<br>Video Models and World Action Modeling</span> </a> <a href="/blog/7d2b8e/" class="timeline-dot" style="left: 81%;"> <span class="timeline-tooltip">May 2026<br>AI for Scientific Discovery</span> </a> <a href="/blog/aw0620/" class="timeline-dot" style="left: 84%;"> <span class="timeline-tooltip">Jun 2026<br>How Qwen-AgentWorld Trains a Language World Model</span> </a> <a href="/blog/gpt-oss-midtraining/" class="timeline-dot" style="left: 86%;"> <span class="timeline-tooltip">Jul 2026<br>Continual Pre-training GPT‑OSS‑120B</span> </a> </div> </div> <div class="news"> <div class="table-responsive"> <table class="table table-sm table-borderless"> <tr> <th scope="row" style="width: 20%">Jul 17, 2026</th> <td> <span class="category-badge badge-llm">llm</span> <a class="news-title" href="/blog/gpt-oss-midtraining/">Continual Pre-training GPT‑OSS‑120B</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Jun 20, 2026</th> <td> <span class="category-badge badge-agent">agent</span> <a class="news-title" href="/blog/aw0620/">How Qwen-AgentWorld Trains a Language World Model</a> </td> </tr> <tr> <th scope="row" style="width: 20%">May 12, 2026</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/7d2b8e/">AI for Scientific Discovery</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Apr 07, 2026</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/w4m407/">Video Models and World Action Modeling</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Mar 11, 2026</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/f1m2r3/">What Does Flow-Matching Bring to Deep RL?</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Feb 15, 2026</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/690518/">Generalizable Value Functions and Introverted Intuition (Ni)</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Jan 09, 2026</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/a7c3e1/">How to Use Privileged Information in RL: On-policy Distillation</a> </td> </tr> <tr class="year-divider"><td colspan="2"><hr></td></tr> <tr> <th scope="row" style="width: 20%">Dec 14, 2025</th> <td> <span class="category-badge badge-llm">llm</span> <a class="news-title" href="/blog/e2b7d4/">Autoregressive Embedding Models: Training, Attention, and Performance</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Nov 22, 2025</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/wx-defense/">Adaptive Sampling and Curriculum Methods</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Oct 01, 2025</th> <td> <span class="category-badge badge-agent">agent</span> <a class="news-title" href="/blog/823eeb/">Position: Why Web is a Good Environment to Study RL?</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Sep 01, 2025</th> <td> <span class="category-badge badge-llm">llm</span> <a class="news-title" href="/blog/19b81f/">Pretraining, Post-training, and Test-Time Reasoning</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Aug 07, 2025</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/a897ea/">Challenges in Scaling Q-Learning</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Jul 22, 2025</th> <td> <span class="category-badge badge-agent">agent</span> <a class="news-title" href="/blog/994498/">Are Multi-step Agents Overthinking?</a> </td> </tr> <tr> <th scope="row" style="width: 20%">May 27, 2025</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/76c93a/">Policy Optimization without a Critic: The GRPO Family</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Mar 15, 2025</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/d01fde/">Can Language Models Be Critic Functions?</a> </td> </tr> <tr class="year-divider"><td colspan="2"><hr></td></tr> <tr> <th scope="row" style="width: 20%">Oct 22, 2024</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/89b69e/">RL on Language under Single-step Settings</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Aug 01, 2024</th> <td> <span class="category-badge badge-llm">llm</span> <a class="news-title" href="/blog/e31c7f/">LLM Optimization Basics: Memory</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Jun 15, 2024</th> <td> <span class="category-badge badge-llm">llm</span> <a class="news-title" href="/blog/9e7104/">LLM Optimization Basics: Time</a> </td> </tr> <tr> <th scope="row" style="width: 20%">May 22, 2024</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/035eca/">Importance Sampling: Why and How</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Apr 07, 2024</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/p1i2t3/">Policy Improvement Theorem</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Mar 13, 2024</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/82d5b0/">The Policy Gradient Family: PG, PPO, and AC</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Feb 18, 2024</th> <td> <span class="category-badge badge-rl">rl</span> <a class="news-title" href="/blog/be11m4n/">Bellman Operator Identities</a> </td> </tr> <tr class="year-divider"><td colspan="2"><hr></td></tr> <tr> <th scope="row" style="width: 20%">Dec 16, 2023</th> <td> <span class="category-badge badge-llm">llm</span> <a class="news-title" href="/blog/m1o2e3/">Mixture of Experts Explained</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Sep 09, 2023</th> <td> <span class="category-badge badge-llm">llm</span> <a class="news-title" href="/blog/9c4e2f/">RoPE and M-RoPE: Rotation, Decay, and Multimodal Axes</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Jun 07, 2023</th> <td> <span class="category-badge badge-llm">llm</span> <a class="news-title" href="/blog/3b5b97/">Self-Attention Layer and The Transformers Architecture</a> </td> </tr> <tr> <th scope="row" style="width: 20%">May 20, 2023</th> <td> <span class="category-badge badge-math">math</span> <a class="news-title" href="/blog/d4p3g1/">Dynamic Programming: Foundations</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Apr 27, 2023</th> <td> <span class="category-badge badge-llm">llm</span> <a class="news-title" href="/blog/b4c5k6/">Backpropagation</a> </td> </tr> </table> </div> </div> <h2 id="talks-blogs">Big Minds</h2> <div class="news"> <div class="table-responsive"> <table class="table table-sm table-borderless"> <tr> <th scope="row" style="width: 20%">Feb 02, 2026</th> <td> <a class="news-title" href="/blog/b7l3cv/">Vincent Sitzmann: The Bitter Lesson of Computer Vision</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Nov 25, 2025</th> <td> <a class="news-title" href="/blog/a2s5r6/">Ilya Sutskever: From the Age of Scaling to the Age of Research</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Aug 15, 2023</th> <td> <a class="news-title" href="/blog/i1s2g3/">Ilya Sutskever: An Observation on Generalization</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Feb 01, 2018</th> <td> <a class="news-title" href="/blog/m3s4p7/">Ilya Sutskever: Meta Learning and Self Play</a> </td> </tr> </table> </div> </div> <h2 id="theory-notes">Theory Notes</h2> <div class="news"> <div class="table-responsive"> <table class="table table-sm table-borderless"> <tr> <th scope="row" style="width: 20%">Jul 09, 2026</th> <td> <span class="category-badge badge-psy">psy</span> <a class="news-title" href="/blog/introverted-intuition/">Introverted Intuition</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Jun 15, 2026</th> <td> <span class="category-badge badge-horo">horo</span> <a class="news-title" href="/blog/h0r012/">Mechanical Watchmaking</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Apr 01, 2026</th> <td> <span class="category-badge badge-psy">psy</span> <a class="news-title" href="/blog/m1b2j3/">Analytical Psychology</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Feb 12, 2026</th> <td> <span class="category-badge badge-music">music</span> <a class="news-title" href="/blog/6840a5/">The Pentatonic Scale</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Dec 13, 2025</th> <td> <span class="category-badge badge-music">music</span> <a class="news-title" href="/blog/dfdef8/">Non-Diatonic Notes</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Sep 18, 2025</th> <td> <span class="category-badge badge-phil">phil</span> <a class="news-title" href="/blog/a1b2c3/">Foundations of Reductionism</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Aug 24, 2025</th> <td> <span class="category-badge badge-music">music</span> <a class="news-title" href="/blog/1da6a5/">Jazz Chords and Their Variants</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Jul 04, 2025</th> <td> <span class="category-badge badge-info">info</span> <a class="news-title" href="/blog/d4e5f6/">Kolmogorov Complexity</a> </td> </tr> <tr> <th scope="row" style="width: 20%">Jun 13, 2025</th> <td> <span class="category-badge badge-music">music</span> <a class="news-title" href="/blog/2c538f/">The Komuro Progression</a> </td> </tr> </table> </div> </div> <h2> <a href="/publications/" style="color: inherit">Selected Publications</a> </h2> <div class="publications"> <ol class="bibliography"> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">arXiv 2026</abbr> </div> <div id="bai2026asyncwebrl" class="col-sm-8"> <div class="title">AsyncWebRL: Efficient Multi-Step RL for Visual Web Agents</div> <div class="author"> <em> <strong>Hao Bai</strong> </em>, Rui Yang, Chenlu Ye, Spencer Whitehead, Aviral Kumar, and Tong Zhang </div> <div class="periodical"> Jun 2026 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://arxiv.org/abs/2606.05597" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> </div> <div class="abstract hidden"> <p>Training vision-language web agents with multi-step reinforcement learning is compute-intensive. We introduce AsyncWebRL, an asynchronous training system that overlaps rollout, gradient update, and policy refresh, together with a trajectory-level objective that removes an efficiency-inducing bias in multi-step GRPO. AsyncWebRL improves end-to-end training throughput and sets a new open-source state of the art on the WebGym out-of-distribution test split.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">CVPR 2026</abbr> </div> <div id="bai2025webgym" class="col-sm-8"> <div class="title">WebGym: Scaling Training Environments for Visual Web Agents with Realistic Tasks</div> <div class="author"> <em> <strong>Hao Bai</strong> </em>, Alexey Taymanov, Tong Zhang, Aviral Kumar, and Spencer Whitehead </div> <div class="periodical"> Jan 2025 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://webgym-website.github.io/webgym.github.io/" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> <a href="https://arxiv.org/pdf/2601.02439" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> <a href="https://github.com/microsoft/webgym" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> <img src="https://img.shields.io/github/stars/microsoft/webgym" alt="GitHub stars"> </div> <div class="abstract hidden"> <p>We present WebGym, the largest-to-date open-source environment for training realistic visual web agents. Real websites are non-stationary and diverse, making artificial or small-scale task sets insufficient for robust policy learning. WebGym contains nearly 300,000 tasks with rubric-based evaluations across diverse, real-world websites and difficulty levels. We train agents with a simple reinforcement learning (RL) recipe, which trains on the agent’s own interaction traces (rollouts), using task rewards as feedback to guide learning. To enable scaling RL, we speed up sampling of trajectories in WebGym by developing a high-throughput asynchronous rollout system, designed specifically for web agents. Our system achieves a 4-5x rollout speedup compared to naive implementations. Second, we scale the task set breadth, depth, and size, which results in continued performance improvement. Fine-tuning a strong base vision-language model, Qwen-3-VL-8B-Instruct, on WebGym results in an improvement in success rate on an out-of-distribution test set from 26.2% to 42.9%, significantly outperforming agents based on proprietary models such as GPT-4o and GPT-5-Thinking that achieve 27.1% and 29.8%, respectively. This improvement is substantial because our test set consists only of tasks on websites never seen during training, unlike many other prior works on training visual web agents.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">NeurIPS 2025</abbr> </div> <div id="shenbai2025tti" class="col-sm-8"> <div class="title">Thinking vs. Doing: Agents that Reason by Scaling Test-Time Interaction</div> <div class="author"> <em> <strong>Hao Bai</strong> </em>, Junhong Shen, Lunjun Zhang, Yifei Zhou, Amrith Setlur, Shengbang Tong, Diego Caples, Nan Jiang, Tong Zhang, Ameet Talwalkar, and Aviral Kumar </div> <div class="periodical"> May 2025 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://test-time-interaction.github.io/" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> <a href="https://arxiv.org/pdf/2506.07976" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> <a href="https://github.com/test-time-interaction/TTI" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> <img src="https://img.shields.io/github/stars/test-time-interaction/TTI" alt="GitHub stars"> </div> <div class="abstract hidden"> <p>Most paradigms for building foundation model agents rely on prompting or finetuning on existing demonstrations, but this is not sufficient in dynamic environments (e.g., mobile device control). In theory, while on-policy reinforcement learning (RL) should address these limitations, this approach itself is not quite effective at leveraging existing agentic data, especially when it is of low quality. An approach to address this issue is to use offline value-based RL but realizing value-based RL for agents has been elusive due to of stability and efficiency associated with running TD-learning at scale with vision-language models (VLMs). In this paper, we develop a scalable value-based RL approach called Digi-Q that makes it possible to train VLM agents with TD-learning. We situate our study in building GUI agents for Android devices. The key idea in Digi-Q is to perform TD-learning on a frozen, intermediate-layer representation of a VLM rather than training the whole VLM itself. Doing so successfully requires an initial phase of fine-tuning to prime VLM representations to feature actionable information that is critical for TD-learning. When done correctly, our approach is able to attain better performance per-unit compute FLOPS. To make maximal use of the learned Q-function, we devise a novel best-of-N policy extraction operator that imitates the best actions out of multiple candidate actions from the current policy as ranked by the value function. With no REINFORCE-style policy gradients that need careful tiuning and an efficient TD-learning approach, Digi-Q outperforms several strong prior methods on user-scale device control tasks in Android-in-the-Wild, attaining 9.9% of relative improvement over prior best-performing offline RL method in this domain.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">ICLR 2025</abbr> </div> <div id="bai2025digiq" class="col-sm-8"> <div class="title">Digi-Q: Transforming VLMs to Device-Control Agents via Value-Based Offline RL</div> <div class="author"> <em> <strong>Hao Bai</strong> </em>, Yifei Zhou, Erran Li, Sergey Levine, and Aviral Kumar </div> <div class="periodical"> Jan 2025 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://digiq-agent.com" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> <a href="https://arxiv.org/pdf/2502.15760" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> <a href="https://github.com/DigiRL-agent/digiq" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> <img src="https://img.shields.io/github/stars/DigiRL-agent/digiq" alt="GitHub stars"> </div> <div class="abstract hidden"> <p>Most paradigms for building foundation model agents rely on prompting or finetuning on existing demonstrations, but this is not sufficient in dynamic environments (e.g., mobile device control). In theory, while on-policy reinforcement learning (RL) should address these limitations, this approach itself is not quite effective at leveraging existing agentic data, especially when it is of low quality. An approach to address this issue is to use offline value-based RL but realizing value-based RL for agents has been elusive due to of stability and efficiency associated with running TD-learning at scale with vision-language models (VLMs). In this paper, we develop a scalable value-based RL approach called Digi-Q that makes it possible to train VLM agents with TD-learning. We situate our study in building GUI agents for Android devices. The key idea in Digi-Q is to perform TD-learning on a frozen, intermediate-layer representation of a VLM rather than training the whole VLM itself. Doing so successfully requires an initial phase of fine-tuning to prime VLM representations to feature actionable information that is critical for TD-learning. When done correctly, our approach is able to attain better performance per-unit compute FLOPS. To make maximal use of the learned Q-function, we devise a novel best-of-N policy extraction operator that imitates the best actions out of multiple candidate actions from the current policy as ranked by the value function. With no REINFORCE-style policy gradients that need careful tiuning and an efficient TD-learning approach, Digi-Q outperforms several strong prior methods on user-scale device control tasks in Android-in-the-Wild, attaining 9.9% of relative improvement over prior best-performing offline RL method in this domain.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">Oral @ CPAL 2025</abbr> </div> <div id="bai2024improving" class="col-sm-8"> <div class="title">Improving Neuron-level Interpretability with White-box Language Models</div> <div class="author"> <em> <strong>Hao Bai</strong> </em>, and Yi Ma </div> <div class="periodical"> Oct 2024 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://crate-lm.github.io/" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> <a href="https://arxiv.org/abs/2410.16443" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> </div> <div class="abstract hidden"> <p>Neurons in auto-regressive language models like GPT-2 can be interpreted by analyzing their activation patterns. Recent studies have shown that techniques such as dictionary learning, a form of post-hoc sparse coding, enhance this neuron-level interpretability. In our research, we are driven by the goal to fundamentally improve neural network interpretability by embedding sparse coding directly within the model architecture, rather than applying it as an afterthought. In our study, we introduce a white-box transformer-like architecture named Coding RAte TransformEr (CRATE), explicitly engineered to capture sparse, low-dimensional structures within data distributions. Our comprehensive experiments showcase significant improvements (up to 103% relative improvement) in neuron-level interpretability across a variety of evaluation metrics. Detailed investigations confirm that this enhanced interpretability is steady across different layers irrespective of the model size, underlining CRATE’s robust performance in enhancing neural network interpretability. Further analysis shows that CRATE’s increased interpretability comes from its enhanced ability to consistently and distinctively activate on relevant tokens. These findings point towards a promising direction for creating white-box foundation models that excel in neuron-level interpretation.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">NeurIPS 2024</abbr> <abbr class="badge rounded w-100"> Oral @ ICML WS</abbr> </div> <div id="bai2024digirl" class="col-sm-8"> <div class="title">DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning</div> <div class="author"> <em> <strong>Hao Bai</strong> </em>, Yifei Zhou, Jiayi Pan, Mert Cemri, Alane Suhr, Sergey Levine, and Aviral Kumar </div> <div class="periodical"> Jun 2024 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://digirl-agent.github.io/" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> <a href="https://arxiv.org/abs/2406.11896" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> <a href="https://github.com/DigiRL-agent/digirl" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> <img src="https://img.shields.io/github/stars/DigiRL-agent/digirl" alt="GitHub stars"> </div> <div class="abstract hidden"> <p>Training corpuses for vision language models typically lack sufficient amounts of decision-centric data. This renders off-the-shelf VLMs sub-optimal for decision-making tasks such as in-the-wild device control through graphical user interfaces (GUIs). While training with static demonstrations has shown some promise, we show that such methods fall short when controlling real GUIs due to their failure to deal with real world stochasticity not captured in static observational data. This paper introduces a novel autonomous RL approach, called DigiRL, for training in-the-wild device control agents through fine-tuning a pre-trained VLM in two stages: offline RL to initialize the model, followed by offline-to-online RL. To do this, we build a scalable and parallelizable Android learning environment equipped with a VLM-based evaluator and develop a simple yet effective RL approach for learning in this domain. Our approach runs advantage-weighted RL with advantage estimators enhanced to account for stochasticity along with an automatic curriculum for deriving maximal learning signal. We demonstrate the effectiveness of DigiRL using the Android-in-the-Wild (AitW) dataset, where our 1.5B VLM trained with RL achieves a 49.5% absolute improvement – from 17.7% to 67.2% success rate – over supervised fine-tuning with static human demonstration data. These results significantly surpass not only the prior best agents, including AppAgent with GPT-4V (8.3% success rate) and the 17B CogAgent trained with AitW data (14.4%), but also the prior best autonomous RL approach based on filtered behavior cloning (57.8%), thereby establishing a new state-of-the-art for digital agents for in-the-wild device control.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">NeurIPS 2024</abbr> </div> <div id="zhai2024fine" class="col-sm-8"> <div class="title">Fine-Tuning Large Vision-Language Models as Decision-Making Agents via Reinforcement Learning</div> <div class="author"> Yuexiang Zhai, <em> <strong>Hao Bai</strong> </em>, Zipeng Lin, Jiayi Pan, Shengbang Tong, Yifei Zhou, Alane Suhr, Saining Xie, Yann LeCun, Yi Ma, and Sergey Levine </div> <div class="periodical"> <em></em> May 2024 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://rl4vlm.github.io/" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> <a href="https://arxiv.org/abs/2405.10292" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> <a href="https://github.com/RL4VLM/RL4VLM" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> <img src="https://img.shields.io/github/stars/RL4VLM/RL4VLM" alt="GitHub stars"> </div> <div class="abstract hidden"> <p>Large vision-language models (VLMs) fine-tuned on specialized visual instruction-following data have exhibited impressive language reasoning capabilities across various scenarios. However, this fine-tuning paradigm may not be able to efficiently learn optimal decision-making agents in multi-step goal-directed tasks from interactive environments. To address this challenge, we propose an algorithmic framework that fine-tunes VLMs with reinforcement learning (RL). Specifically, our framework provides a task description and then prompts the VLM to generate chain-of-thought (CoT) reasoning, enabling the VLM to efficiently explore intermediate reasoning steps that lead to the final text-based action. Next, the open-ended text output is parsed into an executable action to interact with the environment to obtain goal-directed task rewards. Finally, our framework uses these task rewards to fine-tune the entire VLM with RL. Empirically, we demonstrate that our proposed framework enhances the decision-making capabilities of VLM agents across various tasks, enabling 7b models to outperform commercial models such as GPT4-V or Gemini. Furthermore, we find that CoT reasoning is a crucial component for performance improvement, as removing the CoT reasoning results in a significant decrease in the overall performance of our method.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">JMLR</abbr> </div> <div id="yu2023white" class="col-sm-8"> <div class="title">White-Box Transformers via Sparse Rate Reduction: Compression Is All There Is?</div> <div class="author"> Yaodong Yu, Sam Buchanan, Druv Pai, Tianzhe Chu, Ziyang Wu, Shengbang Tong, <em> <strong>Hao Bai</strong> </em>, Yuexiang Zhai, Benjamin D Haeffele, and Yi Ma </div> <div class="periodical"> <em></em> Apr 2024 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://ma-lab-berkeley.github.io/CRATE/" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> <a href="https://arxiv.org/abs/2311.13110" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> <a href="https://github.com/Ma-Lab-Berkeley/CRATE" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> <img src="https://img.shields.io/github/stars/Ma-Lab-Berkeley/CRATE" alt="GitHub stars"> </div> <div class="abstract hidden"> <p>In this paper, we contend that a natural objective of representation learning is to compress and transform the distribution of the data, say sets of tokens, towards a low-dimensional Gaussian mixture supported on incoherent subspaces. The goodness of such a representation can be evaluated by a principled measure, called sparse rate reduction, that simultaneously maximizes the intrinsic information gain and extrinsic sparsity of the learned representation. From this perspective, popular deep network architectures, including transformers, can be viewed as realizing iterative schemes to optimize this measure. Particularly, we derive a transformer block from alternating optimization on parts of this objective: the multi-head self-attention operator compresses the representation by implementing an approximate gradient descent step on the coding rate of the features, and the subsequent multi-layer perceptron sparsifies the features. This leads to a family of white-box transformer-like deep network architectures, named CRATE, which are mathematically fully interpretable. We show, by way of a novel connection between denoising and compression, that the inverse to the aforementioned compressive encoding can be realized by the same class of CRATE architectures. Thus, the so-derived white-box architectures are universal to both encoders and decoders. Experiments show that these networks, despite their simplicity, indeed learn to compress and sparsify representations of large-scale real-world image and text datasets, and achieve performance very close to highly engineered transformer-based models: ViT, MAE, DINO, BERT, and GPT2. We believe the proposed computational framework demonstrates great potential in bridging the gap between theory and practice of deep learning, from a unified perspective of data compression.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">EMNLP’23</abbr> </div> <div id="reddy2023social" class="col-sm-8"> <div class="title">Social Commonsense-Guided Search Query Generation for Open-Domain Knowledge-Powered Conversations</div> <div class="author"> Revanth Reddy, <em> <strong>Hao Bai</strong> </em>, Wentao Yao, Sharath Chandra Etagi Suresh, Heng Ji, and ChengXiang Zhai </div> <div class="periodical"> <em></em> Oct 2023 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://arxiv.org/abs/2310.14340" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> </div> <div class="abstract hidden"> <p>Open-domain dialog involves generating search queries that help obtain relevant knowledge for holding informative conversations. However, it can be challenging to determine what information to retrieve when the user is passive and does not express a clear need or request. To tackle this issue, we present a novel approach that focuses on generating internet search queries that are guided by social commonsense. Specifically, we leverage a commonsense dialog system to establish connections related to the conversation topic, which subsequently guides our query generation. Our proposed framework addresses passive user interactions by integrating topic tracking, commonsense response generation and instructiondriven query generation. Through extensive evaluations, we show that our approach1 overcomes limitations of existing query generation techniques that rely solely on explicit dialog information, and produces search queries that are more relevant, specific, and compelling, ultimately resulting in more engaging responses.</p> </div> </div> </div> </li> </ol> </div> </article> </div> <script>!function(){function t(t){n.textContent=t.getAttribute("data-note"),n.style.opacity="0",n.classList.add("visible");var e=t.getBoundingClientRect(),i=n.getBoundingClientRect(),o=e.left;o+i.width>window.innerWidth-8&&(o=window.innerWidth-i.width-8),o<8&&(o=8),n.style.left=o+"px",n.style.top=e.bottom+4+"px",n.style.opacity=""}function e(){n.classList.remove("visible")}var n=document.createElement("div");n.className="hover-note-tip",document.body.appendChild(n),document.querySelectorAll(".hover-note").forEach(function(i){i.addEventListener("mouseenter",function(){t(i)}),i.addEventListener("mouseleave",e),i.addEventListener("click",function(){n.classList.contains("visible")?e():t(i)})})}();</script> </div> <footer class="sticky-bottom mt-5" role="contentinfo"> <div class="container"> © Copyright 2026 Jack (Hao) Bai. Powered by <a href="https://jekyllrb.com/" target="_blank" rel="external nofollow noopener">Jekyll</a> with <a href="https://github.com/alshedivat/al-folio" rel="external nofollow noopener" target="_blank">al-folio</a> theme. </div> </footer> <script src="https://cdn.jsdelivr.net/npm/jquery@3.6.0/dist/jquery.min.js" integrity="sha256-/xUj+3OJU5yExlq6GSYGSHk7tPXikynS7ogEvDej/m4=" crossorigin="anonymous"></script> <script src="/assets/js/bootstrap.bundle.min.js"></script> <script src="https://cdn.jsdelivr.net/npm/mdbootstrap@4.20.0/js/mdb.min.js" integrity="sha256-NdbiivsvWt7VYCt6hYNT3h/th9vSTL4EDWeGs5SN3DA=" crossorigin="anonymous"></script> <script defer src="https://cdn.jsdelivr.net/npm/masonry-layout@4.2.2/dist/masonry.pkgd.min.js" integrity="sha256-Nn1q/fx0H7SNLZMQ5Hw5JLaTRZp0yILA/FRexe19VdI=" crossorigin="anonymous"></script> <script defer src="https://cdn.jsdelivr.net/npm/imagesloaded@5.0.0/imagesloaded.pkgd.min.js" integrity="sha256-htrLFfZJ6v5udOG+3kNLINIKh2gvoKqwEhHYfTTMICc=" crossorigin="anonymous"></script> <script defer src="/assets/js/masonry.js" type="text/javascript"></script> <script defer src="https://cdn.jsdelivr.net/npm/medium-zoom@1.1.0/dist/medium-zoom.min.js" integrity="sha256-ZgMyDAIYDYGxbcpJcfUnYwNevG/xi9OHKaR/8GK+jWc=" crossorigin="anonymous"></script> <script defer src="/assets/js/zoom.js?3e7054dc4d3e3dd8f0731a48453e618e"></script> <script defer src="/assets/js/sidenote.js?524e5ded3e4d9d041a39f9b243e5e260"></script> <script src="/assets/js/no_defer.js?2781658a0a2b13ed609542042a859126"></script> <script defer src="/assets/js/common.js?0012839f98c37f202100293077bc9852"></script> <script defer src="/assets/js/copy_code.js?3577194613afa04501eb52f8f4164de9" type="text/javascript"></script> <script defer src="/assets/js/jupyter_new_tab.js?d9f17b6adc2311cbabd747f4538bb15f"></script> <script async src="https://d1bxh8uas1mnw7.cloudfront.net/assets/embed.js"></script> <script async src="https://badge.dimensions.ai/badge.js"></script> <style>mjx-container,mjx-container *{-webkit-user-select:text!important;-moz-user-select:text!important;-ms-user-select:text!important;user-select:text!important}.mjx-tex-source{display:inline!important;font-size:0!important;line-height:0!important;color:transparent!important;background:transparent!important;text-shadow:none!important;white-space:pre!important;-webkit-user-select:text!important;user-select:text!important}.mjx-tex-source::selection{color:transparent;background:transparent}.mjx-tex-source::-moz-selection{color:transparent;background:transparent}</style> <script type="text/javascript">!function(){function t(t){if(!window.MathJax||!MathJax.startup||!MathJax.startup.document)return 0;var e=MathJax.startup.document.math,a=0;try{for(var n=Array.from(e),r=0;r<n.length;r++)t(n[r]),a++}catch(n){try{if(e&&e.list&&e.list.next)for(var o=e.list.next;o&&o.data;)t(o.data),a++,o=o.next}catch(t){}}return a}function e(){var e=t(function(t){var e=t.typesetRoot;if(e&&e.setAttribute){var a=t.math||"",n=!!t.display;e.setAttribute("data-tex",a),e.setAttribute("data-display",n?"true":"false");var r=e.querySelector&&e.querySelector(":scope > .mjx-tex-source");r&&r.parentNode.removeChild(r);var o=document.createElement("span");o.className="mjx-tex-source",o.setAttribute("aria-hidden","true"),o.textContent=n?"$$"+a+"$$":"$"+a+"$",e.appendChild(o)}});try{console.log("[mathjax-copy] tagged "+e+" math items")}catch(t){}}function a(){document.addEventListener("copy",function(t){var e=window.getSelection();if(e&&0!==e.rangeCount&&!e.isCollapsed){var a=e.getRangeAt(0);if(a.cloneContents){var n=a.cloneContents();if(n.querySelectorAll){var r=n.querySelectorAll("mjx-container");if(0!==r.length){var o=0;if(r.forEach(function(t){var e=t.getAttribute("data-tex");if(e){var a="true"===t.getAttribute("data-display"),n=document.createTextNode(a?"\n$$"+e+"$$\n":"$"+e+"$");t.parentNode.replaceChild(n,t),o++}}),0!==o){var i=document.createElement("div");i.appendChild(n);var c=i.textContent.replace(/[ \t]+/g," ").replace(/ ?\n ?/g,"\n").replace(/\n{3,}/g,"\n\n").replace(/^\n+|\n+$/g,"");t.clipboardData.setData("text/plain",c),t.preventDefault();try{console.log("[mathjax-copy] intercepted, replaced "+o+" formulas")}catch(t){}}}else{for(var d=a.commonAncestorContainer;d&&1!==d.nodeType;)d=d.parentNode;for(;d&&"MJX-CONTAINER"!==d.tagName;)d=d.parentNode;if(d&&d.getAttribute("data-tex")){var u=d.getAttribute("data-tex")||"",l="true"===d.getAttribute("data-display");return t.clipboardData.setData("text/plain",l?"$$"+u+"$$":"$"+u+"$"),void t.preventDefault()}}}}}})}"loading"===document.readyState?document.addEventListener("DOMContentLoaded",a):a(),window.MathJax={tex:{tags:"ams",inlineMath:[["$","$"],["\\(","\\)"]],displayMath:[["$$","$$"],["\\[","\\]"]],processEscapes:!0},options:{skipHtmlTags:{"[+]":["code","pre","kbd","script","style"]}},startup:{pageReady:function(){return MathJax.startup.defaultPageReady().then(function(){e();var t=MathJax.startup.document;if(t&&"function"==typeof t.updateDocument){var a=t.updateDocument.bind(t);t.updateDocument=function(){var t=a.apply(this,arguments);return e(),t}}})}}}}();</script> <script defer type="text/javascript" id="MathJax-script" src="https://cdn.jsdelivr.net/npm/mathjax@3.2.0/es5/tex-mml-chtml.min.js" integrity="sha256-YyRhy+F3bG+zvvwsW1E/uy5s5pKXwmG/bOPlV+1pt6o=" crossorigin="anonymous"></script> <script defer src="https://cdnjs.cloudflare.com/polyfill/v3/polyfill.min.js?features=es6" crossorigin="anonymous"></script> <script type="text/javascript">function progressBarSetup(){"max"in document.createElement("progress")?(initializeProgressElement(),$(document).on("scroll",function(){progressBar.attr({value:getCurrentScrollPosition()})}),$(window).on("resize",initializeProgressElement)):(resizeProgressBar(),$(document).on("scroll",resizeProgressBar),$(window).on("resize",resizeProgressBar))}function getCurrentScrollPosition(){return $(window).scrollTop()}function initializeProgressElement(){let e=$("#navbar").outerHeight(!0);$("body").css({"padding-top":e}),$("progress-container").css({"padding-top":e}),progressBar.css({top:e}),progressBar.attr({max:getDistanceToScroll(),value:getCurrentScrollPosition()})}function getDistanceToScroll(){return $(document).height()-$(window).height()}function resizeProgressBar(){progressBar.css({width:getWidthPercentage()+"%"})}function getWidthPercentage(){return getCurrentScrollPosition()/getDistanceToScroll()*100}const progressBar=$("#progress");window.onload=function(){setTimeout(progressBarSetup,50)};</script> <script src="/assets/js/vanilla-back-to-top.min.js?982c80efc910ea9a6203400dcbd0a3af"></script> <script>addBackToTop();</script> <script src="/assets/js/shortcut-key.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: 70
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 70 оптимально.
Внутренние ссылки не запрещены к индексации в robots.txt.
Показать внутренние ссылки
| Url | Анкор | Состояние | Анализировать |
|---|---|---|---|
| / |
about <span class="sr-only">(current)</span>
|
|
Анализировать url |
| /assets/pdf/resume.pdf |
here
|
|
|
| /news/ |
News
|
|
Анализировать url |
| /blog/b4c5k6/ |
<span class="timeline-tooltip">Apr 2023<br>Backpropagation</span>
|
|
Анализировать url |
| /blog/3b5b97/ |
<span class="timeline-tooltip">Jun 2023<br>Self-Attention Layer and The Transformers Architecture</span>
|
|
Анализировать url |
| /blog/9c4e2f/ |
<span class="timeline-tooltip">Sep 2023<br>RoPE and M-RoPE: Rotation, Decay, and Multimodal Axes</span>
|
|
Анализировать url |
| /blog/m1o2e3/ |
<span class="timeline-tooltip">Dec 2023<br>Mixture of Experts Explained</span>
|
|
Анализировать url |
| /blog/be11m4n/ |
<span class="timeline-tooltip">Feb 2024<br>Bellman Operator Identities</span>
|
|
Анализировать url |
| /blog/82d5b0/ |
<span class="timeline-tooltip">Mar 2024<br>The Policy Gradient Family: PG, PPO, and AC</span>
|
|
Анализировать url |
| /blog/p1i2t3/ |
<span class="timeline-tooltip">Apr 2024<br>Policy Improvement Theorem</span>
|
|
Анализировать url |
| /blog/035eca/ |
<span class="timeline-tooltip">May 2024<br>Importance Sampling: Why and How</span>
|
|
Анализировать url |
| /blog/9e7104/ |
<span class="timeline-tooltip">Jun 2024<br>LLM Optimization Basics: Time</span>
|
|
Анализировать url |
| /blog/e31c7f/ |
<span class="timeline-tooltip">Aug 2024<br>LLM Optimization Basics: Memory</span>
|
|
Анализировать url |
| /blog/89b69e/ |
<span class="timeline-tooltip">Oct 2024<br>RL on Language under Single-step Settings</span>
|
|
Анализировать url |
| /blog/d01fde/ |
<span class="timeline-tooltip">Mar 2025<br>Can Language Models Be Critic Functions?</span>
|
|
Анализировать url |
| /blog/76c93a/ |
<span class="timeline-tooltip">May 2025<br>Policy Optimization without a Critic: The GRPO Family</span>
|
|
Анализировать url |
| /blog/994498/ |
<span class="timeline-tooltip">Jul 2025<br>Are Multi-step Agents Overthinking?</span>
|
|
Анализировать url |
| /blog/a897ea/ |
<span class="timeline-tooltip">Aug 2025<br>Challenges in Scaling Q-Learning</span>
|
|
Анализировать url |
| /blog/19b81f/ |
<span class="timeline-tooltip">Sep 2025<br>Pretraining, Post-training, and Test-Time Reasoning</span>
|
|
Анализировать url |
| /blog/823eeb/ |
<span class="timeline-tooltip">Oct 2025<br>Position: Why Web is a Good Environment to Study RL?</span>
|
|
Анализировать url |
| /blog/wx-defense/ |
<span class="timeline-tooltip">Nov 2025<br>Adaptive Sampling and Curriculum Methods</span>
|
|
Анализировать url |
| /blog/e2b7d4/ |
<span class="timeline-tooltip">Dec 2025<br>Autoregressive Embedding Models: Training, Attention, and Performance</span>
|
|
Анализировать url |
| /blog/a7c3e1/ |
<span class="timeline-tooltip">Jan 2026<br>How to Use Privileged Information in RL: On-policy Distillation</span>
|
|
Анализировать url |
| /blog/690518/ |
<span class="timeline-tooltip">Feb 2026<br>Generalizable Value Functions and Introverted Intuition (Ni)</span>
|
|
Анализировать url |
| /blog/f1m2r3/ |
<span class="timeline-tooltip">Mar 2026<br>What Does Flow-Matching Bring to Deep RL?</span>
|
|
Анализировать url |
| /blog/w4m407/ |
<span class="timeline-tooltip">Apr 2026<br>Video Models and World Action Modeling</span>
|
|
Анализировать url |
| /blog/7d2b8e/ |
<span class="timeline-tooltip">May 2026<br>AI for Scientific Discovery</span>
|
|
Анализировать url |
| /blog/aw0620/ |
<span class="timeline-tooltip">Jun 2026<br>How Qwen-AgentWorld Trains a Language World Model</span>
|
|
Анализировать url |
| /blog/gpt-oss-midtraining/ |
<span class="timeline-tooltip">Jul 2026<br>Continual Pre-training GPT‑OSS‑120B</span>
|
|
Анализировать url |
| /blog/gpt-oss-midtraining/ |
Continual Pre-training GPT‑OSS‑120B
|
|
Анализировать url |
| /blog/aw0620/ |
How Qwen-AgentWorld Trains a Language World Model
|
|
Анализировать url |
| /blog/7d2b8e/ |
AI for Scientific Discovery
|
|
Анализировать url |
| /blog/w4m407/ |
Video Models and World Action Modeling
|
|
Анализировать url |
| /blog/f1m2r3/ |
What Does Flow-Matching Bring to Deep RL?
|
|
Анализировать url |
| /blog/690518/ |
Generalizable Value Functions and Introverted Intuition (Ni)
|
|
Анализировать url |
| /blog/a7c3e1/ |
How to Use Privileged Information in RL: On-policy Distillation
|
|
Анализировать url |
| /blog/e2b7d4/ |
Autoregressive Embedding Models: Training, Attention, and Performance
|
|
Анализировать url |
| /blog/wx-defense/ |
Adaptive Sampling and Curriculum Methods
|
|
Анализировать url |
| /blog/823eeb/ |
Position: Why Web is a Good Environment to Study RL?
|
|
Анализировать url |
| /blog/19b81f/ |
Pretraining, Post-training, and Test-Time Reasoning
|
|
Анализировать url |
| /blog/a897ea/ |
Challenges in Scaling Q-Learning
|
|
Анализировать url |
| /blog/994498/ |
Are Multi-step Agents Overthinking?
|
|
Анализировать url |
| /blog/76c93a/ |
Policy Optimization without a Critic: The GRPO Family
|
|
Анализировать url |
| /blog/d01fde/ |
Can Language Models Be Critic Functions?
|
|
Анализировать url |
| /blog/89b69e/ |
RL on Language under Single-step Settings
|
|
Анализировать url |
| /blog/e31c7f/ |
LLM Optimization Basics: Memory
|
|
Анализировать url |
| /blog/9e7104/ |
LLM Optimization Basics: Time
|
|
Анализировать url |
| /blog/035eca/ |
Importance Sampling: Why and How
|
|
Анализировать url |
| /blog/p1i2t3/ |
Policy Improvement Theorem
|
|
Анализировать url |
| /blog/82d5b0/ |
The Policy Gradient Family: PG, PPO, and AC
|
|
Анализировать url |
| /blog/be11m4n/ |
Bellman Operator Identities
|
|
Анализировать url |
| /blog/m1o2e3/ |
Mixture of Experts Explained
|
|
Анализировать url |
| /blog/9c4e2f/ |
RoPE and M-RoPE: Rotation, Decay, and Multimodal Axes
|
|
Анализировать url |
| /blog/3b5b97/ |
Self-Attention Layer and The Transformers Architecture
|
|
Анализировать url |
| /blog/d4p3g1/ |
Dynamic Programming: Foundations
|
|
Анализировать url |
| /blog/b4c5k6/ |
Backpropagation
|
|
Анализировать url |
| /blog/b7l3cv/ |
Vincent Sitzmann: The Bitter Lesson of Computer Vision
|
|
Анализировать url |
| /blog/a2s5r6/ |
Ilya Sutskever: From the Age of Scaling to the Age of Research
|
|
Анализировать url |
| /blog/i1s2g3/ |
Ilya Sutskever: An Observation on Generalization
|
|
Анализировать url |
| /blog/m3s4p7/ |
Ilya Sutskever: Meta Learning and Self Play
|
|
Анализировать url |
| /blog/introverted-intuition/ |
Introverted Intuition
|
|
Анализировать url |
| /blog/h0r012/ |
Mechanical Watchmaking
|
|
Анализировать url |
| /blog/m1b2j3/ |
Analytical Psychology
|
|
Анализировать url |
| /blog/6840a5/ |
The Pentatonic Scale
|
|
Анализировать url |
| /blog/dfdef8/ |
Non-Diatonic Notes
|
|
Анализировать url |
| /blog/a1b2c3/ |
Foundations of Reductionism
|
|
Анализировать url |
| /blog/1da6a5/ |
Jazz Chords and Their Variants
|
|
Анализировать url |
| /blog/d4e5f6/ |
Kolmogorov Complexity
|
|
Анализировать url |
| /blog/2c538f/ |
The Komuro Progression
|
|
Анализировать url |
| /publications/ |
Selected Publications
|
|
Анализировать url |
Внешние ссылки
Кол-во: 37
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 37 слишком много. Спрячьте лишние ссылки в тег noindex или атрибут rel='nofollow'!
На странице ссылки с атрибутом rel='nofollow' 18.
Показать внешние ссылки
| Url | Анкор | Анализировать |
|---|---|---|
| scholar.google.com |
<i class="ai ai-google-scholar"></i>
|
Анализировать url |
| github.com |
<i class="fa-brands fa-github"></i>
|
Анализировать url |
| linkedin.com |
<i class="fa-brands fa-linkedin"></i>
|
Анализировать url |
| twitter.com |
<i class="fa-brands fa-x-twitter"></i>
|
|
| zhihu.com |
<i class="fa-brands fa-zhihu"></i>
|
Анализировать url |
| tongzhang-ml.org |
Tong Zhang
|
Анализировать url |
| aviralkumar2907.github.io |
Aviral Kumar
|
Анализировать url |
| nanjiang.cs.illinois.edu |
Nan Jiang
|
Анализировать url |
| people.eecs.berkeley.edu |
Sergey Levine
|
Анализировать url |
| arxiv.org |
arXiv
|
Анализировать url |
| arxiv.org |
ArXiv
|
Анализировать url |
| webgym-website.github.io |
project page
|
Анализировать url |
| arxiv.org |
ArXiv
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| webgym-website.github.io |
HTML
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| test-time-interaction.github.io |
HTML
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| digiq-agent.com |
HTML
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| crate-lm.github.io |
HTML
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| digirl-agent.github.io |
HTML
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| rl4vlm.github.io |
HTML
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| ma-lab-berkeley.github.io |
HTML
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| github.com |
Code
|
Анализировать url |
| arxiv.org |
PDF
|
Анализировать url |
| jekyllrb.com |
Jekyll
|
Анализировать url |
| github.com |
al-folio
|
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
Оптимизация сайта jackgethome.com выполнена всего на 40%. Чтобы исправить ситуацию и попасть в ТОП необходимо:
Постарайтесь исправить общие ошибки.
Исправьте ошибки в мета-тегах.
Исправьте ошибки индексации.
Поделитесь с друзьями: