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> MedGRPO: Multi-Task Reinforcement Learning for Heterogeneous Medical Video Understanding | Yuhao Su </title> <meta name="author" content="Yuhao Su"> <meta name="description" content="A 531K-pair medical video benchmark (MedVidBench) covering 8 tasks, plus a multi-task RL post-training framework whose cross-dataset reward normalization keeps training stable when heterogeneous tasks are mixed. MedVidBench is the benchmark of the ECCV 2026 MedVidU Challenge."> <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" href="/assets/css/scholar-icons.css?62b2ac103a88034e6882a5be5f3e2772"> <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="data:image/svg+xml,<svg%20xmlns=%22http://www.w3.org/2000/svg%22%20viewBox=%220%200%20100%20100%22><text%20y=%22.9em%22%20font-size=%2290%22>%F0%9F%8E%93</text></svg>"> <link rel="stylesheet" href="/assets/css/main.css?d41d8cd98f00b204e9800998ecf8427e"> <link rel="canonical" href="https://yuhaosu.github.io/projects/medgrpo/">
1<script src="/assets/js/theme.js?a81d82887dd692e91686b43de4542f18"></script>
1 <link defer rel="stylesheet" href="/assets/css/jekyll-pygments-themes-native.css?5847e5ed4a4568527aa6cfab446049ca" media="none" id="highlight_theme_dark">
1<script> 2 initTheme(); 3 </script>
3 </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="/"> Yuhao Su </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 "> <a class="nav-link" href="/#news"> News </a> </li> <li class="nav-item "> <a class="nav-link" href="/#selected-works"> Selected Works </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">MedGRPO: Multi-Task Reinforcement Learning for Heterogeneous Medical Video Understanding</h1> <p class="post-meta"><strong>CVPR 2026</strong></p> <p class="post-description">A 531K-pair medical video benchmark (MedVidBench) covering 8 tasks, plus a multi-task RL post-training framework whose cross-dataset reward normalization keeps training stable when heterogeneous tasks are mixed. MedVidBench is the benchmark of the ECCV 2026 MedVidU Challenge.</p> </header> <article> <p>MedGRPO introduces a comprehensive framework for medical video understanding, combining a large-scale dataset with a novel multi-task reinforcement learning approach.</p> <h2 id="medvidbench-dataset">MedVidBench Dataset</h2> <ul> <li> <strong>531K video-instruction pairs</strong>: Large-scale dataset for medical video understanding</li> <li> <strong>8 diverse tasks</strong>: Covers video-level, segment-level, and frame-level understanding</li> <li>Enables comprehensive evaluation of medical video models</li> </ul> <h2 id="medgrpo-framework">MedGRPO Framework</h2> <ul> <li> <strong>Multi-Task Reinforcement Learning</strong>: Novel RL framework designed for heterogeneous medical video tasks</li> <li> <strong>Cross-Dataset Reward Normalization</strong>: Ensures balanced training across different tasks and datasets</li> <li> <strong>Medical LLM Judge</strong>: Specialized evaluation for medical video caption quality</li> </ul> <h2 id="performance-improvements">Performance Improvements</h2> <p>Compared to supervised fine-tuning (SFT) baseline:</p> <ul> <li> <strong>+0.074 [email protected]</strong> on temporal action grounding</li> <li> <strong>+0.588 LLM score</strong> on video summary generation</li> <li>Consistent improvements across all 8 tasks</li> </ul> </article> </div> </div> <footer class="fixed-bottom" role="contentinfo"> <div class="container mt-0"> © Copyright 2026 Yuhao Su. © 2025 Yuhao Su. Powered by <a href="https://jekyllrb.com/" target="_blank" rel="external nofollow noopener">Jekyll</a> with the <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>. </div> </footer>
3<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/jquery.min.js" integrity="sha256-/xUj+3OJU5yExlq6GSYGSHk7tPXikynS7ogEvDej/m4=" crossorigin="anonymous"></script>
3
3<script src="/assets/js/bootstrap.bundle.min.js"></script>
3
3<script src="https://cdn.jsdelivr.net/npm/[email protected]/js/mdb.min.js" integrity="sha256-NdbiivsvWt7VYCt6hYNT3h/th9vSTL4EDWeGs5SN3DA=" crossorigin="anonymous"></script>
3
3<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/masonry.pkgd.min.js" integrity="sha256-Nn1q/fx0H7SNLZMQ5Hw5JLaTRZp0yILA/FRexe19VdI=" crossorigin="anonymous"></script>
3
3<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/imagesloaded.pkgd.min.js" integrity="sha256-htrLFfZJ6v5udOG+3kNLINIKh2gvoKqwEhHYfTTMICc=" crossorigin="anonymous"></script>
3
3<script defer src="/assets/js/masonry.js?a0db7e5d5c70cc3252b3138b0c91dcaf" type="text/javascript"></script>
3
3<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/medium-zoom.min.js" integrity="sha256-ZgMyDAIYDYGxbcpJcfUnYwNevG/xi9OHKaR/8GK+jWc=" crossorigin="anonymous"></script>
3
3<script defer src="/assets/js/zoom.js?85ddb88934d28b74e78031fd54cf8308"></script>
3
3<script src="/assets/js/no_defer.js?2781658a0a2b13ed609542042a859126"></script>
3
3<script defer src="/assets/js/common.js?c15de51d4bb57887caa2c21988d97279"></script>
3
3<script defer src="/assets/js/copy_code.js?c8a01c11a92744d44b093fc3bda915df" type="text/javascript"></script>
3
3<script defer src="/assets/js/jupyter_new_tab.js?d9f17b6adc2311cbabd747f4538bb15f"></script>
3
3<script async src="https://d1bxh8uas1mnw7.cloudfront.net/assets/embed.js"></script>
3
3<script async src="https://badge.dimensions.ai/badge.js"></script>
3
3<script defer type="text/javascript" id="MathJax-script" src="https://cdn.jsdelivr.net/npm/[email protected]/es5/tex-mml-chtml.js" integrity="sha256-MASABpB4tYktI2Oitl4t+78w/lyA+D7b/s9GEP0JOGI=" crossorigin="anonymous"></script>
3
3<script src="/assets/js/mathjax-setup.js?a5bb4e6a542c546dd929b24b8b236dfd"></script>
3
3<script defer src="https://cdnjs.cloudflare.com/polyfill/v3/polyfill.min.js?features=es6" crossorigin="anonymous"></script>
3
3<script defer src="/assets/js/progress-bar.js?2f30e0e6801ea8f5036fa66e1ab0a71a" type="text/javascript"></script>
3
3<script src="/assets/js/vanilla-back-to-top.min.js?f40d453793ff4f64e238e420181a1d17"></script>
3
3<script> 4 addBackToTop(); 5 </script>
5 </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.