PageSourceSearch

https://zhoujiayi003.github.io/publications/

html zhoujiayi003.github.io collected 2026-10-03 10:14:46 UTC 21,955 bytes, 11 lines download raw bytes

1<!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> publications | Jiayi Zhou </title> <meta name="author" content="Jiayi Zhou"> <meta name="description" content="grouped by authorship role. &lt;sup&gt;*&lt;/sup&gt; denotes equal contribution."> <meta name="keywords" content="jekyll, jekyll-theme, academic-website, portfolio-website"> <link rel="stylesheet" href="/assets/css/bootstrap.min.css?a4b3f509e79c54a512b890d73235ef04"> <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/css/mdb.min.css" integrity="sha256-jpjYvU3G3N6nrrBwXJoVEYI/0zw8htfFnhT9ljN3JJw=" crossorigin="anonymous"> <link defer rel="stylesheet" href="/assets/css/academicons.min.css?f0b7046b84e425c55f3463ac249818f5"> <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&amp;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="data:image/svg+xml,&lt;svg%20xmlns=%22http://www.w3.org/2000/svg%22%20viewBox=%220%200%20100%20100%22&gt;&lt;text%20y=%22.9em%22%20font-size=%2290%22&gt;%E2%9A%9B%EF%B8%8F&lt;/text&gt;&lt;/svg&gt;"> <link rel="stylesheet" href="/assets/css/main.css?d41d8cd98f00b204e9800998ecf8427e"> <link rel="canonical" href="https://zhoujiayi003.github.io/publications/"> 
1<script src="/assets/js/theme.js?9a0c749ec5240d9cda97bc72359a72c0"></script>
1 <link defer rel="stylesheet" href="/assets/css/jekyll-pygments-themes-native.css?5847e5ed4a4568527aa6cfab446049ca" media="none" id="highlight_theme_dark"> 
1<script>initTheme();</script>
1 </head> <body class="fixed-top-nav "> <header> <nav id="navbar" class="navbar navbar-light navbar-expand-sm fixed-top" role="navigation"> <div class="container"> <a class="navbar-brand title font-weight-lighter" href="/"> <span class="font-weight-bold">Jiayi</span> Zhou </a> <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 "> <a class="nav-link" href="/">about </a> </li> <li class="nav-item active"> <a class="nav-link" href="/publications/">publications <span class="sr-only">(current)</span> </a> </li> <li class="nav-item "> <a class="nav-link" href="/cv/">cv </a> </li> <li class="nav-item"> <button id="search-toggle" title="Search" onclick="openSearchModal()"> <span class="nav-link">ctrl k <i class="ti ti-search"></i></span> </button> </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">publications</h1> <p class="post-description">grouped by authorship role. <sup>*</sup> denotes equal contribution.</p> </header> <article> <div class="publications"> <h2>First author / co-first author</h2> <h2 class="bibliography">2026</h2> <ol class="bibliography"><li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100"> <a href="https://neurips.cc/" rel="external nofollow noopener" target="_blank">NeurIPS</a> </abbr> </div> <div id="he2026videoodyssey" class="col-sm-8"> <div class="title">VideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding</div> <div class="author"> Haichen He<sup>*</sup> , <em>Jiayi Zhou<sup>*</sup></em>, Sifeng Shang , and <span class="more-authors" title="click to view 3 more authors" onclick=" var element=$(this); element.attr('title', ''); var more_authors_text=element.text() == '3 more authors' ? 'Yihan Hu, Yuanhan Zhang, Kaiyang Zhou' : '3 more authors'; var cursorPosition=0; var textAdder=setInterval(function(){ element.html(more_authors_text.substring(0, cursorPosition + 1)); if (++cursorPosition == more_authors_text.length){ clearInterval(textAdder); } }, '10'); ">3 more authors</span> <i class="fa-solid fa-circle-info ml-1" data-toggle="popover" data-placement="top" data-html="true" data-content="* Equal contribution"> </i> </div> <div class="periodical"> <em>In Advances in Neural Information Processing Systems (NeurIPS), Evaluations &amp; Datasets Track</em> , 2026 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="http://arxiv.org/abs/2605.22907" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">arXiv</a> <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> <a href="https://github.com/maifoundations/VideoOdyssey" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> <a href="https://videoodyssey-project.github.io/" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Website</a> </div> <div class="abstract hidden"> <p>Real-world long video understanding requires models to perform continuous tracking, information integration and memory retention over massive temporal spans within extreme video durations. Mastering this intense cognitive load constitutes the fundamental bottleneck in long video understanding. While existing benchmarks have driven progress by scaling up video duration, their evaluation tasks often require comprehending only short and isolated video segments, falling short of capturing the challenge of ultra-long-context reasoning. To measure this cognitive load, we emphasize continuous certificate length, defined as the video length a human must continuously watch to definitively answer a given question. Driven by this metric, we introduce VideoOdyssey, a benchmark specifically designed for ultra-long-context and omni-modal video understanding. VideoOdyssey is characterized by three key features: 1) Extreme video duration and diversity: spanning 11 domains and 54 subcategories with an average video duration of 109 minutes; 2) Comprehensive evaluation scenarios: offering two subsets to address different research focuses, i.e., VideoOdyssey-V for probing the limits of visual understanding in MLLMs, and VideoOdyssey-AV for evaluating synchronized audio-visual understanding for omni-modal models; 3) Ultra-long and multi-level continuous certificates: extending the average continuous certificate to 16 minutes for VideoOdyssey-V and 12.8 minutes for VideoOdyssey-AV. Crucially, we design 5 granular levels from seconds to hours, providing a comprehensive diagnostic tool to evaluate models across varying context lengths and cognitive loads. Extensive evaluations show that bottlenecks of current MLLMs extend beyond simple retrieval to include struggles with continuous reasoning across varying context lengths, fine-grained perception, and non-verbal omni-modal understanding. We hope VideoOdyssey will spur the development of next-generation MLLMs toward genuine real-world video understanding.</p> </div> <div class="bibtex hidden"> <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">he2026videoodyssey</span><span class="p">,</span>
2  <span class="na">title</span> <span class="p">=</span> <span class="s">{VideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding}</span><span class="p">,</span>
3  <span class="na">author</span> <span class="p">=</span> <span class="s">
3{He, Haichen and Zhou, Jiayi and Shang, Sifeng and Hu, Yihan and Zhang, Yuanhan and Zhou, Kaiyang}</span><span class="p">,</span>
4  <span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Advances in Neural Information Processing Systems (NeurIPS), Evaluations \&amp; Datasets Track}</span><span class="p">,</span>
5  <span class="na">year</span> <span class="p">=</span> <span class="s">{2026}</span><span class="p">,</span>
6<span class="p">}</span></code></pre></figure> </div> </div> </div> </li></ol> <h2>Second author / other</h2> <h2 class="bibliography">2026</h2> <ol class="bibliography"><li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100"> <a href="https://iclr.cc/" rel="external nofollow noopener" target="_blank">ICLR</a> </abbr> </div> <div id="shang2026qzo" class="col-sm-8"> <div class="title">Fine-tuning Quantized Neural Networks with Zeroth-order Optimization</div> <div class="author"> Sifeng Shang , <em>Jiayi Zhou</em>, Chenyu Lin , and <span class="more-authors" title="click to view 2 more authors" onclick=" var element=$(this); element.attr('title', ''); var more_authors_text=element.text() == '2 more authors' ? 'Minxian Li, Kaiyang Zhou' : '2 more authors'; var cursorPosition=0; var textAdder=setInterval(function(){ element.html(more_authors_text.substring(0, cursorPosition + 1)); if (++cursorPosition == more_authors_text.length){ clearInterval(textAdder); } }, '10'); ">2 more authors</span> </div> <div class="periodical"> <em>In International Conference on Learning Representations (ICLR)</em> , 2026 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="http://arxiv.org/abs/2505.13430" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">arXiv</a> <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> <a href="https://openreview.net/forum?id=mGeeRFToaW" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> <a href="https://github.com/maifoundations/QZO" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> </div> <div class="abstract hidden"> <p>As the size of large language models grows exponentially, GPU memory has become a bottleneck for adapting these models to downstream tasks. In this paper, we aim to push the limits of memory-efficient training by minimizing memory usage on model weights, gradients, and optimizer states, within a unified framework. Our idea is to eliminate both gradients and optimizer states using zeroth-order optimization, which approximates gradients by perturbing weights during forward passes to identify gradient directions. To minimize memory usage on weights, we employ model quantization, e.g., converting from bfloat16 to int4. However, directly applying zeroth-order optimization to quantized weights is infeasible due to the precision gap between discrete weights and continuous gradients, which would otherwise require de-quantization and re-quantization. To overcome this challenge, we propose Quantized Zeroth-order Optimization (QZO), a simple yet effective approach that perturbs the continuous quantization scale for gradient estimation and uses a directional derivative clipping method to stabilize training. QZO is orthogonal to both scalar-based and codebook-based post-training quantization methods. Compared to full-parameter fine-tuning in 16 bits, QZO can reduce the total memory cost by more than 18x for 4-bit LLMs, and enables fine-tuning Llama-2-13B within a single 24GB GPU.</p> </div> <div class="bibtex hidden"> <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">shang2026qzo</span><span class="p">,</span>
7  <span class="na">title</span> <span class="p">=</span> <span class="s">{Fine-tuning Quantized Neural Networks with Zeroth-order Optimization}</span><span class="p">,</span>
8  <span class="na">author</span> <span class="p">=</span> <span class="s">{Shang, Sifeng and Zhou, Jiayi and Lin, Chenyu and Li, Minxian and Zhou, Kaiyang}</span><span class="p">,</span>
9  <span class="na">booktitle</span> <span class="p">=</span> <span class="s">{International Conference on Learning Representations (ICLR)}</span><span class="p">,</span>
10  <span class="na">year</span> <span class="p">=</span> <span class="s">{2026}</span><span class="p">,</span>
11<span class="p">}</span></code></pre></figure> </div> </div> </div> </li></ol> </div> </article> </div> </div> <footer class="fixed-bottom" role="contentinfo"> <div class="container mt-0"> © Copyright 2026 Jiayi Zhou. 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. Hosted by <a href="https://pages.github.com/" target="_blank" rel="external nofollow noopener">GitHub Pages</a>. Photos from <a href="https://unsplash.com" target="_blank" rel="external nofollow noopener">Unsplash</a>. </div> </footer> 
11<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/jquery.min.js" integrity="sha256-/xUj+3OJU5yExlq6GSYGSHk7tPXikynS7ogEvDej/m4=" crossorigin="anonymous"></script>
11 
11<script src="/assets/js/bootstrap.bundle.min.js"></script>
11 
11<script src="https://cdn.jsdelivr.net/npm/[email protected]/js/mdb.min.js" integrity="sha256-NdbiivsvWt7VYCt6hYNT3h/th9vSTL4EDWeGs5SN3DA=" crossorigin="anonymous"></script>
11 
11<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/masonry.pkgd.min.js" integrity="sha256-Nn1q/fx0H7SNLZMQ5Hw5JLaTRZp0yILA/FRexe19VdI=" crossorigin="anonymous"></script>
11 
11<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/imagesloaded.pkgd.min.js" integrity="sha256-htrLFfZJ6v5udOG+3kNLINIKh2gvoKqwEhHYfTTMICc=" crossorigin="anonymous"></script>
11 
11<script defer src="/assets/js/masonry.js" type="text/javascript"></script>
11 
11<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/medium-zoom.min.js" integrity="sha256-ZgMyDAIYDYGxbcpJcfUnYwNevG/xi9OHKaR/8GK+jWc=" crossorigin="anonymous"></script>
11 
11<script defer src="/assets/js/zoom.js?85ddb88934d28b74e78031fd54cf8308"></script>
11 
11<script src="/assets/js/no_defer.js?2781658a0a2b13ed609542042a859126"></script>
11 
11<script defer src="/assets/js/common.js?e0514a05c5c95ac1a93a8dfd5249b92e"></script>
11 
11<script defer src="/assets/js/copy_code.js?12775fdf7f95e901d7119054556e495f" type="text/javascript"></script>
11 
11<script defer src="/assets/js/jupyter_new_tab.js?d9f17b6adc2311cbabd747f4538bb15f"></script>
11 
11<script async src="https://d1bxh8uas1mnw7.cloudfront.net/assets/embed.js"></script>
11 
11<script async src="https://badge.dimensions.ai/badge.js"></script>
11 
11<script type="text/javascript">window.MathJax={tex:{tags:"ams"}};</script>
11 
11<script defer type="text/javascript" id="MathJax-script" src="https://cdn.jsdelivr.net/npm/[email protected]/es5/tex-mml-chtml.min.js" integrity="sha256-rjmgmaB99riUNcdlrDtcAiwtLIojSxNyUFdl+Qh+rB4=" crossorigin="anonymous"></script>
11 
11<script defer src="https://cdnjs.cloudflare.com/polyfill/v3/polyfill.min.js?features=es6" crossorigin="anonymous"></script>
11 
11<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=$("#progress");window.onload=function(){setTimeout(progressBarSetup,50)};</script>
11 
11<script src="/assets/js/vanilla-back-to-top.min.js?f40d453793ff4f64e238e420181a1d17"></script>
11 
11<script>addBackToTop();</script>
11 
11<script type="module" src="/assets/js/search/ninja-keys.min.js?601a2d3465e2a52bec38b600518d5f70"></script>
11 <ninja-keys hidebreadcrumbs noautoloadmdicons placeholder="Type to start searching"></ninja-keys> 
11<script>let searchTheme=determineComputedTheme();const ninjaKeys=document.querySelector("ninja-keys");"dark"===searchTheme?ninjaKeys.classList.add("dark"):ninjaKeys.classList.remove("dark");const openSearchModal=()=>{const e=$("#navbarNav");e.hasClass("show")&&e.collapse("hide"),ninjaKeys.open()};</script>
11 
11<script>const ninja=document.querySelector("ninja-keys");ninja.data=[{id:"nav-about",title:"about",section:"Navigation",handler:()=>{window.location.href="/"}},{id:"nav-publications",title:"publications",description:"grouped by authorship role. * denotes equal contribution.",section:"Navigation",handler:()=>{window.location.href="/publications/"}},{id:"nav-cv",title:"cv",description:"",section:"Navigation",handler:()=>{window.location.href="/cv/"}},{id:"news-our-paper-lt-a-href-quot-https-arxiv-org-abs-2605-22907-quot-gt-videoodyssey-a-benchmark-for-ultra-long-context-and-omni-modal-video-understanding-lt-a-gt-has-been-accepted-to-lt-strong-gt-neurips-2026-lt-strong-gt-evaluations-amp-amp-datasets-track-tada",title:'Our paper <a href="https://arxiv.org/abs/2605.22907">VideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding</a> has been accepted to <strong>NeurIPS 2026</strong> (Evaluations &amp; Datasets Track)! <img class="emoji" title=":tada:" alt=":tada:" src="https://github.githubassets.com/images/icons/emoji/unicode/1f389.png" height="20" width="20">',description:"",section:"News"},{id:"news-our-paper-lt-a-href-quot-https-arxiv-org-abs-2505-13430-quot-gt-qzo-fine-tuning-quantized-neural-networks-with-zeroth-order-optimization-lt-a-gt-has-been-accepted-to-lt-strong-gt-iclr-2026-lt-strong-gt-tada",title:'Our paper <a href="https://arxiv.org/abs/2505.13430">QZO: Fine-tuning Quantized Neural Networks with Zeroth-order Optimization</a> has been accepted to <strong>ICLR 2026</strong>! <img class="emoji" title=":tada:" alt=":tada:" src="https://github.githubassets.com/images/icons/emoji/unicode/1f389.png" height="20" width="20">',description:"",section:"News"},{id:"news-create-my-first-homepage-sparkles-smile",title:'Create my first homepage! <img class="emoji" title=":sparkles:" alt=":sparkles:" src="https://github.githubassets.com/images/icons/emoji/unicode/2728.png" height="20" width="20"> <img class="emoji" title=":smile:" alt=":smile:" src="https://github.githubassets.com/images/icons/emoji/unicode/1f604.png" height="20" width="20">',description:"",section:"News"},{id:"projects-project-1",title:"project 1",description:"with background image",section:"Projects",handler:()=>{window.location.href="/projects/1_project/"}},{id:"projects-project-2",title:"project 2",description:"a project with a background image and giscus comments",section:"Projects",handler:()=>{window.location.href="/projects/2_project/"}},{id:"projects-project-3-with-very-long-name",title:"project 3 with very long name",description:"a project that redirects to another website",section:"Projects",handler:()=>{window.location.href="/projects/3_project/"}},{id:"projects-project-4",title:"project 4",description:"another without an image",section:"Projects",handler:()=>{window.location.href="/projects/4_project/"}},{id:"projects-project-5",title:"project 5",description:"a project with a background image",section:"Projects",handler:()=>{window.location.href="/projects/5_project/"}},{id:"projects-project-6",title:"project 6",description:"a project with no image",section:"Projects",handler:()=>{window.location.href="/projects/6_project/"}},{id:"projects-project-7",title:"project 7",description:"with background image",section:"Projects",handler:()=>{window.location.href="/projects/7_project/"}},{id:"projects-project-8",title:"project 8",description:"an other project with a background image and giscus comments",section:"Projects",handler:()=>{window.location.href="/projects/8_project/"}},{id:"projects-project-9",title:"project 9",description:"another project with an image \ud83c\udf89",section:"Projects",handler:()=>{window.location.href="/projects/9_project/"}},{id:"socials-email",title:"Send email",section:"Socials",handler:()=>{window.open("mailto:%7A%68%6F%75%6A%69%61%79%69%30%30%33 %61%74 %67%6D%61%69%6C.%63%6F%6D","_blank")}},{id:"socials-google-scholar",title:"Google Scholar",section:"Socials",handler:()=>{window.open("https://scholar.google.com/citations?user=d2NhPNMAAAAJ","_blank")}},{id:"light-theme",title:"Change theme to light",description:"Change the theme of the site to Light",section:"Theme",handler:()=>{setThemeSetting("light")}},{id:"dark-theme",title:"Change theme to dark",description:"Change the theme of the site to Dark",section:"Theme",handler:()=>{setThemeSetting("dark")}},{id:"system-theme",title:"Use system default theme",description:"Change the theme of the site to System Default",section:"Theme",handler:()=>{setThemeSetting("system")}}];</script>
11 
11<script src="/assets/js/shortcut-key.js"></script>
11 </body> </html>

Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.