1<!DOCTYPE html> <html lang="en" data-language="en" data-default-language=""> <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>Projects | Freedom AI</title> <meta name="author" content="Freedom AI "> <meta name="description" content="A simple, whitespace theme for academics. Based on [*folio](https://github.com/bogoli/-folio) design. "> <link href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" rel="stylesheet" integrity="sha256-DF7Zhf293AJxJNTmh5zhoYYIMs2oXitRfBjY+9L//AY=" crossorigin="anonymous"> <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/css/mdb.min.css" integrity="sha256-jpjYvU3G3N6nrrBwXJoVEYI/0zw8htfFnhT9ljN3JJw=" crossorigin="anonymous"> <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/@fortawesome/[email protected]/css/all.min.css" integrity="sha256-mUZM63G8m73Mcidfrv5E+Y61y7a12O5mW4ezU3bxqW4=" crossorigin="anonymous"> <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/css/academicons.min.css" integrity="sha256-i1+4qU2G2860dGGIOJscdC30s9beBXjFfzjWLjBRsBg=" crossorigin="anonymous"> <link 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"> <link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jwarby/jekyll-pygments-themes@master/github.css" media="none" id="highlight_theme_light"> <link rel="shortcut icon" href="/assets/img/icon.png">
1<script src="/assets/js/block_polyfill_io.js"></script>
1
1<script src="/assets/js/wechat_popover.js?v=2" defer></script>
1 <link rel="stylesheet" href="/assets/css/main.css?v=1790064493"> <link rel="canonical" href="https://hkunlp.github.io/projects/"> <link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jwarby/jekyll-pygments-themes@master/native.css" media="none" id="highlight_theme_dark">
1<script src="/assets/js/theme.js"></script>
1
1<script src="/assets/js/dark_mode.js"></script>
1 </head> <body class="fixed-top-nav sticky-bottom-footer"> <link rel="stylesheet" href="/assets/css/site-navigation.css?v=1790064493"> <header> <nav id="navbar" class="navbar navbar-light navbar-expand-lg fixed-top"> <div class="container"> <a class="navbar-brand title font-weight-lighter" href="/">Freedom AI</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="/"><span data-lang="en">About</span><span data-lang="zh">å
³äº</span></a> </li> <li class="nav-item "> <a class="nav-link" href="/people/"><span data-lang="en">People</span><span data-lang="zh">å¢é</span></a> </li> <li class="nav-item "> <a class="nav-link" href="/publications/"><span data-lang="en">Publications</span><span data-lang="zh">论æ</span></a> </li> <li class="nav-item active"> <a class="nav-link" href="/projects/"><span data-lang="en">Projects</span><span data-lang="zh">项ç®</span><span class="sr-only">(current)</span></a> </li> <li class="nav-item "> <a class="nav-link" href="/life/"><span data-lang="en">Life</span><span data-lang="zh">çæ´»</span></a> </li> <li class="nav-item "> <a class="nav-link" href="/impact/"><span data-lang="en">Impact</span><span data-lang="zh">å½±åå</span></a> </li> <li class="nav-item "> <a class="nav-link" href="/startups/"><span data-lang="en">Startups</span><span data-lang="zh">åä¸</span></a> </li> <li class="nav-item "> <a class="nav-link" href="/blog/"><span data-lang="en">Blog</span><span data-lang="zh">å客</span></a> </li> <li class="nav-item "> <a class="nav-link" href="/join-us/"><span data-lang="en">Join Us</span><span data-lang="zh">å å
¥æä»¬</span></a> </li> </ul> </div> <div class="site-nav-actions"> <button id="light-toggle" type="button" title="Change theme" aria-label="Change theme"> <i class="fas fa-moon" aria-hidden="true"></i> <i class="fas fa-sun" aria-hidden="true"></i> </button> <button id="language-toggle" type="button" title="Switch language" aria-label="Switch language"><span data-language-label>䏿</span></button> </div> </div> </nav> </header> <div class="container mt-5"> <div class="post"> <header class="post-header"> <h1 class="post-title"> <span data-lang="en">Projects</span><span data-lang="zh">项ç®</span> </h1> <p class="post-description"></p> </header><article> <link rel="stylesheet" href="/assets/css/project-hub.css?v=1790064493"> <div class="project-hub project-focused"> <div class="project-filter project-filter-toolbar" aria-label="Project filters"> <button type="button" class="is-active" data-project-filter="all"><span data-lang="en">All</span><span data-lang="zh">å
¨é¨</span></button> <button type="button" data-project-filter="Foundation Models"><span data-lang="en">Foundation Models</span><span data-lang="zh">åºç¡å¤§æ¨¡å</span></button> <button type="button" data-project-filter="Reasoning & Agents"><span data-lang="en">Reasoning & Agents</span><span data-lang="zh">æ¨ç䏿ºè½ä½</span></button> <button type="button" data-project-filter="AI Systems & Infrastructure"><span data-lang="en">AI Systems & Infrastructure</span><span data-lang="zh">AI ç³»ç»ä¸åºç¡è®¾æ½</span></button> <button type="button" data-project-filter="World Models & Simulation"><span data-lang="en">World Models & Simulation</span><span data-lang="zh">ä¸ç模å䏿¨¡æ</span></button> <button type="button" data-project-filter="Healthcare AI"><span data-lang="en">Healthcare AI</span><span data-lang="zh">å»ç人工æºè½</span></button> <button type="button" data-project-filter="Legal AI"><span data-lang="en">Legal AI</span><span data-lang="zh">æ³å¾äººå·¥æºè½</span></button> <button type="button" data-project-filter="AI Education"><span data-lang="en">AI Education</span><span data-lang="zh">AI æè²</span></button> <button type="button" data-project-filter="Evaluation & Interpretability"><span data-lang="en">Evaluation & Interpretability</span><span data-lang="zh">è¯æµä¸å¯è§£éæ§</span></button> </div> <p class="project-empty" hidden=""><span data-lang="en">No projects found for this tag.</span><span data-lang="zh">该æ ç¾ä¸ææ 项ç®ã</span></p> <section class="project-program-section project-theme-multimodal" id="multimodal-llms" data-project-tags="Foundation Models;Reasoning & Agents;AI Systems &
1 Infrastructure;Healthcare AI" data-project-url="/projects/multimodal-foundation-models/"> <div class="project-program-heading"> <div class="project-eyebrow">Multimodal foundation models and visual reasoning</div> <h2>夿¨¡æå¤§æ¨¡å</h2> <p> 夿¨¡æå¤§æ¨¡åæ¹åææåãå¾åãè§é¢ãé³é¢åå»å¦è§è§æ¾è¿åä¸ä¸ªè½åå°å¾ï¼ä» LongLLaVA/MileBench çé¿ä¸ä¸æè§è§çè§£ï¼å° TRIM çè§è§ token å缩ãShareGPT-4o-Image/Janus-4o ç弿¾å¾åçæï¼åå° Video-R1ãHuatuoGPT-Vision å FusionAudio è¿ç±»é¢åæ¨çãå»çåé³é¢åºæ¯ç模å䏿°æ®ã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/multimodal-foundation-models/">Open Multimodal LLMs</a> <a class="project-action" href="/projects/long-context-multimodal/">Open LongLLaVA and MileBench</a> <a class="project-action" href="/projects/sharegpt-4o-image-janus-4o/">Open ShareGPT-4o and Janus-4o</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/media-roundup-2026/longllava-architecture.webp" alt="Long-context multimodal LLM architecture"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2409.02889" target="_blank" rel="noopener">LongLLaVA: Scaling Multi-modal LLMs to 1000 Images Efficiently</a>ï¼é¿ä¸ä¸æå¤å¾åçè§£åæ¨çã</li> <li> <a href="https://arxiv.org/abs/2404.18532" target="_blank" rel="noopener">MileBench: Benchmarking MLLMs in Long Context</a>ï¼å¤å¾åãè§é¢åé¿è§è§ä¸ä¸æè¯æµã</li> <li> <a href="https://arxiv.org/abs/2409.10994" target="_blank" rel="noopener">TRIM: Less is More for Efficient Multi-modal LLMs</a>ï¼åå°åä½è§è§ tokenï¼éä½å¤æ¨¡ææ¨çææ¬ã</li> <li> <a href="https://arxiv.org/abs/2506.18095" target="_blank" rel="noopener">ShareGPT-4o-Image and Janus-4o</a>ï¼å¼æ¾ GPT-4o 飿 ¼å¾åçæ/ç¼è¾æ°æ®åç»ä¸å¤æ¨¡ææ¨¡åã</li> <li> <a href="https://arxiv.org/abs/2503.21776" target="_blank" rel="noopener">Video-R1: Reinforcing Video Reasoning in MLLMs</a>ï¼æ R1-style reasoning training æ©å±å°è§é¢çè§£ã</li> <li> <a href="https://arxiv.org/abs/2406.19280" target="_blank" rel="noopener">HuatuoGPT-Vision</a>ï¼é¢åå»å¦è§è§ç¥è¯æ³¨å
¥ç夿¨¡æå»å¦å¤§æ¨¡åã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>模å / æ°æ® / 代ç å
¥å£</span><strong>Multimodal stack</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="/projects/multimodal-foundation-models/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>夿¨¡æå¤§æ¨¡å</strong><span>æ»è§è§è§çè§£ãè§è§çæãè§é¢æ¨çãå»å¦è§è§åé³é¢å¤æ¨¡æã</span></span></a> <a class="project-stack-row" href="/projects/long-context-multimodal/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>LongLLaVA and MileBench</strong><span>é¿ä¸ä¸æå¤æ¨¡ææ¨¡åå benchmarkã</span></span></a> <a class="project-stack-row" href="/projects/sharegpt-4o-image-janus-4o/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>ShareGPT-4o-Image and Janus-4o</strong><span>弿¾å¾åçæ/ç¼è¾æ°æ®åç»ä¸å¤æ¨¡ææ¨¡åã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/LongLLaVA" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/LongLLaVA</strong><span>é¿ä¸ä¸æå¤æ¨¡ææ¨¡åèµæºã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/TRIM" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/TRIM</strong><span>髿夿¨¡æ token reductionã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/HuatuoGPT-Vision" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>HuatuoGPT-Vision</strong><span>å»å¦è§è§è¯è¨æ¨¡ååæ°æ®èµæºã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-infra" id="llm-efficiency-ai-infra" data-project-tags="AI Systems &
1 Infrastructure;Reasoning & Agents" data-project-url="/projects/llm-efficiency-ai-infra/"> <div class="project-program-heading"> <div class="project-eyebrow">Efficient training, inference, retrieval, and multimodal context</div> <h2>LLM efficiency and AI Infra</h2> <p> è¿æ¡çº¿æâè½ä¸è½è·å¾èµ·ãè®å¾å¨ãæ¹å¾å¿«âä½ä¸ºé¡¹ç®ä¸»é¢ï¼è¦çæ¨çè·¯å¾åªæãquestion-free fine-tuningãprefix fine-tuningãeditable efficient RAGãé¿ä¸ä¸æå¤æ¨¡æå token reductionã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/llm-efficiency-ai-infra/">Open LLM Infra</a> <a class="project-action" href="/projects/long-context-multimodal/">Open LongLLaVA and MileBench</a> <a class="project-action" href="/projects/rag-instruction-data/">Open Data Infrastructure</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/media-roundup-2026/longllava-architecture.webp" alt="LLM efficiency and AI infrastructure"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2604.16029" target="_blank" rel="noopener">Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoning</a>ï¼æååªæä½ä»·å¼æ¨çè·¯å¾ã</li> <li> <a href="https://arxiv.org/abs/2506.12860" target="_blank" rel="noopener">Question-Free Fine-Tuning: Towards Efficient and Adaptive Reasoning in Large Language Models</a>ï¼é¢å髿èªéåºæ¨çç QFFTã</li> <li> <a href="https://arxiv.org/abs/2503.02875" target="_blank" rel="noopener">The First Few Tokens Are All You Need</a>ï¼UPFTï¼ç¨æå°åç¼ token è¿è¡é«ææ çç£æ¨çå¾®è°ã</li> <li> <a href="https://papers.cool/venue/COLM.2025" target="_blank" rel="noopener">E2-RAG: Towards Editable Efficient RAG by Editing Compressed KV Caches</a>ï¼ç¼è¾å缩 KV cacheï¼é¢åå¿«éæ´æ°åºæ¯ã</li> <li> <a href="https://arxiv.org/abs/2409.02889" target="_blank" rel="noopener">LongLLaVA: Scaling Multi-modal LLMs to 1000 Images Efficiently</a>ï¼é¿ä¸ä¸æå¤å¾å夿¨¡ææ¨çã</li> <li> <a href="https://arxiv.org/abs/2409.10994" target="_blank" rel="noopener">TRIM: Less is More for Efficient Multi-modal LLMs</a>ï¼åå°è§è§ tokenï¼éä½å¤æ¨¡ææ¨çææ¬ã</li> </ol> </div> </div> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="/projects/native-ai-compute-adaptation/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong><span data-lang="en">China-Developed AI Computing</span><span data-lang="zh">å½äº§ AI ç®åéé
ä¸å é</span></strong><span><span data-lang="en">Ascend training, full-stack acceleration, benchmark validation, real-world delivery, and open ecosystem building.</span><span data-lang="zh">æè
¾å¤§è§æ¨¡è®ç»ãå
¨æ 计ç®å éãBenchmark éªè¯ãçå®åºæ¯äº¤ä»ä¸å¼æ¾çæã</span></span></span></a> </div> <details class="project-resources-panel"> <summary><span>GitHub / èµæºå
¥å£</span><strong>Infrastructure stack</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/LongLLaVA" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/LongLLaVA</strong><span>é¿ä¸ä¸æå¤æ¨¡ææ¨¡åã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/TRIM" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/TRIM</strong><span>髿夿¨¡æ token reductionã</span></span></a> <a class="project-stack-row" href="https://github.com/1KE-JI/UPFT" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>1KE-JI/UPFT</strong><span>Unsupervised prefix fine-tuning resourcesã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/RAG-Instruct" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/RAG-Instruct</strong><span>Retrieval-augmented instruction dataã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/LLMZoo" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/LLMZoo</strong><span>弿¾æ¨¡ååè®ç»èµæºéåã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-infra" id="native-ai-compute" data-project-tags="AI Systems &
1 Infrastructure" data-project-url="/projects/native-ai-compute-adaptation/"> <div class="project-program-heading"> <div class="project-eyebrow">China-developed accelerators, inference systems, and open benchmarks</div> <h2> <span data-lang="en">China-Developed AI Computing: Adaptation and Acceleration</span><span data-lang="zh">å½äº§ AI ç®åéé
ä¸å é</span> </h2> <p><span data-lang="en">End-to-end engineering for large-scale Ascend training, model migration, vLLM/SGLang and KV Cache optimization, operators and kernels, benchmark validation, real-world delivery, and open ecosystem building.</span><span data-lang="zh">è¦çæè
¾å¤§è§æ¨¡è®ç»ã模åè¿ç§»ãvLLM/SGLang ä¸ KV Cache ä¼åãç®åä¸ KernelãBenchmark éªè¯ãçå®åºæ¯äº¤ä»å弿¾çæå»ºè®¾çå
¨æ ææ¯é¾æ¡ã</span></p> <div class="project-actions"> <a class="project-action primary" href="/projects/native-ai-compute-adaptation/">Open Compute Adaptation Project</a> <a class="project-action" href="https://github.com/FreedomIntelligence/AccelMark" target="_blank" rel="noopener">Open AccelMark</a> <a class="project-action" href="https://huggingface.co/FreedomIntelligence/openPangu-Embedded-7B" target="_blank" rel="noopener">openPangu on Ascend</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"><img src="/assets/img/native-compute-stack.svg" alt="China-developed AI compute adaptation and acceleration stack"></div> <div> <h3> <span data-lang="en">From accelerator to application</span><span data-lang="zh">ä»ç®åå°åºç¨</span> </h3> <ol class="project-paper-list"> <li> <strong>Scale:</strong> Arabic LLM training on 2,368 Ascend 910A processors, with results at NAACL 2024, ACL 2025, and NeurIPS 2024.</li> <li> <strong>Systems:</strong> prompt, KV Cache, inference-engine, attention, operator, and kernel optimization.</li> <li> <strong>Evaluation:</strong> CANNBench results and AccelMark's transparent, reproducible heterogeneous-accelerator comparisons.</li> <li> <strong>Delivery:</strong> localized medical agents, HuatuoGPT, and intelligent health-management workflows in a secure hospital environment.</li> <li> <strong>Ecosystem:</strong> 100+ Hugging Face models, million-scale downloads, and 10,000+ GitHub Stars across representative projects.</li> </ol> </div> </div> </section> <section class="project-program-section project-theme-optimization" id="llm-reasoning-agentic-rl" data-project-tags="Reasoning & Agents;AI Systems & Infrastructure" data-project-url="/projects/llm-reasoning-agentic-rl/"> <div class="project-program-heading"> <div class="project-eyebrow">Reasoning, RL, and adaptive test-time compute</div> <h2>LLM Reasoning & Agentic RL</h2> <p> This project organizes the lab's recent work on verifiable reasoning, policy optimization, path pruning, code-integrated thinking, and multimodal R1-style training. The through-line is simple: make LLMs reason with feedback signals that are inspectable, efficient, and useful for downstream agents. </p> <div class="project-actions"> <a class="project-action primary" href="/projects/llm-reasoning-agentic-rl/">Open LLM Reasoning & Agentic RL</a> <a class="project-action" href="/projects/reasoning-math-optimization/">Open Math and Optimization</a> <a class="project-action" href="https://wabyking.github.io/" target="_blank" rel="noopener">Papers from Benyou Wang</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/freedomai-2026/impact/iclr-financial-ai-award.jpg" alt="LLM reasoning and agentic reinforcement learning project signal"> </div> <div> <h3>Paper organization</h3> <ol class="project-paper-list"> <li> <a href="https://wabyking.github.io/" target="_blank" rel="noopener">OnePO: Direct One-stage Policy Optimization for SFT-free Domain Adaptation</a> - direct policy optimization without a separate SFT stage.</li> <li> <a href="https://arxiv.org/abs/2605.25511" target="_blank" rel="noopener">CRPO: Character-centric Group Relative Policy Optimization for Role-aware Reasoning in Role-playing Agents</a> - RL objectives for role-aware reasoning agents.</li> <li> <a href="https://arxiv.org/abs/2506.12860" target="_blank" rel="noopener">Question-Free Fine-Tuning</a>
1 - efficient and adaptive reasoning fine-tuning.</li> <li> <a href="https://arxiv.org/abs/2604.16029" target="_blank" rel="noopener">Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoning</a> - learnable path pruning for large reasoning models.</li> <li> <a href="https://arxiv.org/abs/2503.21776" target="_blank" rel="noopener">Video-R1: Reinforcing Video Reasoning in MLLMs</a> - R1-style reinforcement learning for multimodal video reasoning.</li> <li> <a href="https://arxiv.org/abs/2506.09820" target="_blank" rel="noopener">CoRT: Code-integrated Reasoning within Thinking</a> - executable computation inside the reasoning process.</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>Project stack</span><strong>Reasoning papers, code, and datasets</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="https://bijiaxihh.github.io/STOP/" target="_blank" rel="noopener"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>STOP / Cut Your Losses</strong><span>Early path pruning for efficient parallel reasoning.</span></span></a> <a class="project-stack-row" href="https://github.com/tulerfeng/Video-R1" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>Video-R1</strong><span>Video-R1-CoT and RL training resources for video reasoning.</span></span></a> <a class="project-stack-row" href="https://github.com/ChengpengLi1003/CoRT" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>CoRT</strong><span>Code-integrated reasoning resources.</span></span></a> <a class="project-stack-row" href="https://github.com/1KE-JI/UPFT" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>UPFT</strong><span>Unsupervised prefix fine-tuning for efficient reasoning models.</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-human" id="llm-agents-applications" data-project-tags="Reasoning & Agents;World Models & Simulation" data-project-url="/projects/llm-agents-applications/"> <div class="project-program-heading"> <div class="project-eyebrow">Agents, applications, simulators, and human-agent interaction</div> <h2>LLM Agents, Applications & Human-Agent Interaction</h2> <p> This project groups papers where LLMs become agents and interactive systems: tool planners, user simulators, standardized patients, role-playing agents, market participants, speech-to-speech systems, and micro-world actors. The goal is to organize agent papers by what the agent does, what environment it acts in, how it interacts with humans, and how the interaction is evaluated. </p> <div class="project-actions"> <a class="project-action primary" href="/projects/llm-agents-applications/">Open Agent Applications</a> <a class="project-action" href="/projects/human-agent-interaction/">Open Interaction Details</a> <a class="project-action" href="https://wabyking.github.io/" target="_blank" rel="noopener">Papers from Benyou Wang</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/media-roundup-2026/twinmarket-bdi.png" alt="LLM agents and applications project map"> </div> <div> <h3>Paper organization</h3> <ol class="project-paper-list"> <li> <a href="https://aclanthology.org/2025.naacl-long.169/" target="_blank" rel="noopener">Smurfs: Multi-Agent System using Context-Efficient DFSDT for Tool Planning</a> - multi-agent tool planning with context-efficient search.</li> <li> <a href="https://arxiv.org/abs/2308.11534" target="_blank" rel="noopener">Large Language Model as a User Simulator</a> - LLM users for dialogue training and evaluation.</li> <li> <a href="https://openreview.net/forum?id=9nddtu94uX" target="_blank" rel="noopener">PlatoLM: Teaching LLMs via a Socratic Questioning User Simulator</a> - Socratic interaction as a training signal.</li> <li> <a href="https://arxiv.org/abs/2511.14783" target="_blank" rel="noopener">Human or LLM as Standardized Patients?</a> - AI patients for medical education and evaluation.</li> <li> <a href="https://arxiv.org/abs/2602.24080" target="_blank" rel="noopener">Human or Machine? A Preliminary Turing Test for Speech-to-Speech Interaction</a> - human-likeness evaluation for speech-to-speech interaction.</li> <li> <a href="https://arxiv.org/abs/2502.01506" target="_blank" rel="noopener">TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets</a> - LLM investor agents in market simulation.</li> <li> <a href="https://arxiv.org/abs/2603.00585" target="_blank" rel="noopener">MicroVerse</a> - agentic micro-world simulation for scientific processes.</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>Project stack</span><strong>Agent applications and environments</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="https://arxiv.org/abs/2405.05955" target="_blank" rel="noopener"><span class="project-stack-type">Paper</span><span class="project-stack-main"><strong>Smurfs</strong><span>
1Multi-agent tool planning through context-efficient DFSDT.</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/EasyMED" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>EasyMED</strong><span>AI standardized patient framework and evaluation resources.</span></span></a> <a class="project-stack-row" href="/projects/twinmarket-financial-simulation/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>TwinMarket</strong><span>Financial market simulation with LLM agents.</span></span></a> <a class="project-stack-row" href="/projects/microverse-world-simulation/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>MicroVerse</strong><span>Micro-world simulation with hidden mechanisms and evolving states.</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-environment" id="environment-engineering-world-models" data-project-tags="World Models & Simulation;Reasoning & Agents" data-project-url="/projects/environment-engineering-world-models/"> <div class="project-program-heading"> <div class="project-eyebrow">Environment engineering and world models</div> <h2>ç¯å¢å·¥ç¨åä¸ç模å</h2> <p> è¿æ¡çº¿æâ模åè½åâæ¾è¿å¯è¿è¡ç¯å¢ä¸ï¼å®ä¹ç¶æãè¡å¨ãå·¥å
·ãç¨æ·ãè§åãä¸çå¨åå¦åè¯æµå¯¹é½ï¼è®©æºè½ä½è½å¨ç¨æ·æ¨¡æãéèå¸åºãç§å¦å¾®ä¸çå坿§è¡ä¼åç¯å¢ä¸è®ç»ãè¯æµåè¿ä»£ã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/environment-engineering-world-models/">Open Environment Engineering</a> <a class="project-action" href="https://wabyking.github.io/" target="_blank" rel="noopener">Benyou Wang Papers</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/economic-world-models/ewm-hero.png" alt="Environment engineering connects agents, environments, co-evolution, and real-world alignment"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2603.00585" target="_blank" rel="noopener">MicroVerse: A Preliminary Exploration Toward a Micro-World Simulation</a>ï¼æ world model æ©å±å°å¨å®ãç»èåäºç»èè¿ç¨ï¼å¼ºè°éèæºå¶ãç¶ææ¼ååç§å¦çº¦æã</li> <li> <a href="https://arxiv.org/abs/2502.01506" target="_blank" rel="noopener">TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets</a>ï¼ç¨ LLM æèµè
æé éèå¸åºç¯å¢ï¼ä»ä¸ªä½ä¿¡å¿µåä¿¡æ¯æµçæå®è§è¡ä¸ºã</li> <li> <a href="/assets/pdf/economic-world-models.pdf" target="_blank" rel="noopener">From Economic Agents to Agentic Economies</a>ï¼ææºè½ä½ãç¯å¢ãå
±æ¼ååçå®ä¸ç对é½ç»ç»æç»æµä¸ç模åçç³»ç»èå¾ã</li> <li> <a href="https://arxiv.org/abs/2308.11534" target="_blank" rel="noopener">Large Language Model as a User Simulator</a>ï¼ç¨ LLM æé 坿§ç¨æ·ç¯å¢ï¼ä¸ºå¤è½®å¯¹è¯è®ç»åè¯æµæä¾å¯éå¤äº¤äºå¯¹è±¡ã</li> <li> <a href="https://arxiv.org/abs/2405.17743" target="_blank" rel="noopener">ORLM: A Customizable Framework in Training Large Models for Automated Optimization Modeling</a>ï¼æèªç¶è¯è¨é®é¢è½¬æä¼å模åã约æãç®æ 彿°å坿§è¡ä»£ç ã</li> <li> <a href="https://arxiv.org/abs/2510.04204" target="_blank" rel="noopener">CALM Before the STORM</a>ï¼ç ç©¶ä¼å建模ä¸çåçæ¨çè½åï¼æå·¥ç¨é®é¢ç»ç»æå¯éªè¯å»ºæ¨¡ç¯å¢ã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>é¡¹ç® / 论æ / 代ç èµæº</span><strong>Environment stack</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="/projects/environment-engineering-world-models/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>ç¯å¢å·¥ç¨åä¸ç模å</strong><span>ç»ç»ç¯å¢å®ä¹ãæºè½ä½æ¥å£ãä¸çå¨åå¦åè¯æµå¯¹é½çæ»è§é¡µã</span></span></a> <a class="project-stack-row" href="/projects/economic-world-models/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>Economic World Models</strong><span>ç»æµä¸çæ¨¡åæ¶æãè½åé¶æ¢¯å agentic economiesã</span></span></a> <a class="project-stack-row" href="/projects/twinmarket-financial-simulation/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>TwinMarket</strong><span>éèå¸åºä¸ç夿ºè½ä½ç¤¾ä¼ä»¿çç¯å¢ã</span></span></a> <a class="project-stack-row" href="/projects/microverse-world-simulation/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>MicroVerse</strong><span>é¢åçç©åç§å¦æºå¶çå¾®ä¸ç模æã</span></span></a> <a class="project-stack-row" href="https://github.com/Cardinal-Operations/ORLM" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>Cardinal-Operations/ORLM</strong><span>èªå¨åä¼å建模è®ç»æ¡æ¶å benchmarkã</span></span></a> <a class="project-stack-row" href="https://github.com/tangzhy/STORM" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>tangzhy/STORM</strong><span>CALM/STORM ä¼å建模æ¨çèµæºã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-human" id="mobile-agent" data-project-tags="Reasoning & Agents;AI Systems &
1 Infrastructure" data-project-url="/projects/mobile-agent/"> <div class="project-program-heading"> <div class="project-eyebrow">Phone-use agents and mobile workflow execution</div> <h2>Mobile Agent</h2> <p> è¿æ¡çº¿æææºåæéç¨æºè½ä½çç宿§è¡åºï¼ä»éç§ä¸å®å
¨è¯æµï¼å°å¯æ§ææºç¯å¢ãGUI / CLI / tool mixed-action harnessï¼åå°é¢åçå® App workflow ç open model è®ç»ï¼è®© phone-use agents è½å¯é ãå®å
¨ãå¯å®¡è®¡å°å®æä»»å¡ã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/mobile-agent/">Open Mobile Agent</a> <a class="project-action" href="https://phoneharness.github.io/" target="_blank" rel="noopener">Open PhoneHarness</a> <a class="project-action" href="https://huggingface.co/datasets/PhoneHarness/phoneharness-bench" target="_blank" rel="noopener">Open Dataset</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/media-roundup-2026/phoneharness-overview.png" alt="PhoneHarness mobile agent overview"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2604.00986" target="_blank" rel="noopener">Do Phone-Use Agents Respect Your Privacy?</a>ï¼ç¨ MyPhoneBench è¯æµææºæºè½ä½çæéãæ«é²åè®°å¿æ§å¶é£é©ã</li> <li> <a href="https://arxiv.org/abs/2605.07630" target="_blank" rel="noopener">Safe, or Simply Incapable?</a>ï¼åºåçæ£å®å
¨ãå±é©è¡å¨åâä»ä¹ä¹åä¸å¥½âç失败ã</li> <li> <a href="https://arxiv.org/abs/2605.29486" target="_blank" rel="noopener">PhoneWorld</a>ï¼æçå® GUI è½¨è¿¹åæªå¾è½¬æå¯å¤ç¨ã坿§è¡ãå¯éªè¯çææºç¯å¢ã</li> <li> <a href="https://arxiv.org/abs/2606.14832" target="_blank" rel="noopener">PhoneHarness</a>ï¼ç¨ GUIãCLI å tool action çæ··åå¨ä½ç©ºé´è¯æµç宿æºå·¥ä½æµå®æåº¦ã</li> <li> <a href="https://arxiv.org/abs/2606.23049" target="_blank" rel="noopener">PhoneBuddy</a>ï¼ç»åçå® App ç¯å¢å mock phone environment è®ç»å¼æ¾ phone-use agent 模åã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>é¡¹ç® / 代ç / æ°æ®èµæº</span><strong>Mobile agent stack</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="/projects/mobile-agent/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>Mobile Agent</strong><span>ææºæºè½ä½çè¯æµãå®å
¨ãç¯å¢ãharness åè®ç»è·¯çº¿æ»è§ã</span></span></a> <a class="project-stack-row" href="https://phoneharness.github.io/" target="_blank" rel="noopener"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>PhoneHarness</strong><span>Mixed GUIãCLI å tool action çææºæºè½ä½é¡¹ç®ä¸»é¡µã</span></span></a> <a class="project-stack-row" href="https://github.com/PhoneHarness/PhoneHarness" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>PhoneHarness/PhoneHarness</strong><span>Phone-use agent harness å evaluation workflow 代ç ã</span></span></a> <a class="project-stack-row" href="https://huggingface.co/datasets/PhoneHarness/phoneharness-bench" target="_blank" rel="noopener"><span class="project-stack-type">HF Dataset</span><span class="project-stack-main"><strong>PhoneHarness Bench</strong><span>ææºæºè½ä½ mixed-action benchmark æ°æ®èµæºã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-optimization" id="agentic-creation-executable-artifacts" data-project-tags="Reasoning & Agents;World Models & Simulation" data-project-url="/projects/agentic-creation-executable-artifacts/"> <div class="project-program-heading"> <div class="project-eyebrow">From instructions to runnable games, CAD scripts, solvers, and code</div> <h2>Agentic Creation & Executable Artifacts</h2> <p> è¿æ¡çº¿å
³æ³¨ç䏿¯â模å说å¾åä¸åâï¼èæ¯æºè½ä½è½ä¸è½æèªç¶è¯è¨éæ±åæçæ£å¯è¿è¡ã坿¸²æãå¯éªè¯ç产ç©ã代表任å¡å
æ¬çæå¯ç©ç Godot 游æã坿§è¡ç Blender CAD èæ¬ãæ°å¦å»ºæ¨¡æ±è§£å¨ï¼ä»¥åææ¨çè¿ç¨æ¥å°ä»£ç æ§è¡ç¯å¢ä¸ã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/agentic-creation-executable-artifacts/">Open Agentic Creation</a> <a class="project-action" href="https://github.com/FreedomIntelligence/gamecraft-bench" target="_blank" rel="noopener">Open GameCraft-Bench</a> <a class="project-action" href="https://github.com/FreedomIntelligence/BlenderLLM" target="_blank" rel="noopener">Open BlenderLLM</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/freedomai-2026/impact/github-repos.png" alt="Executable artifacts from agentic creation"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2606.17861" target="_blank" rel="noopener">GameCraft-Bench</a>ï¼è¯ä¼° coding agents æ¯å¦è½ææ¸¸æè§æ ¼çæå®æ´ãå¯å¯å¨ãå¯ replay ç Godot 项ç®ã</li> <li> <a href="https://arxiv.org/abs/2412.14203" target="_blank" rel="noopener">BlenderLLM</a>
1ï¼è®ç» LLM çæå¯å¨ Blender 䏿§è¡å¹¶æ¸²æ 3D 模åç CAD èæ¬ã</li> <li> <a href="https://arxiv.org/abs/2405.13144" target="_blank" rel="noopener">Mamo</a>ï¼é¢åæ°å¦å»ºæ¨¡ç benchmark å solverï¼æèªç¶è¯è¨é®é¢è½¬æå¯æ±è§£æ¨¡åã</li> <li> <a href="https://arxiv.org/abs/2405.17743" target="_blank" rel="noopener">ORLM</a>ï¼è®ç»å¤§æ¨¡ååèªå¨åä¼å建模ï¼è¿æ¥é®é¢æè¿°ã约æãç®æ 彿°å代ç ã</li> <li> <a href="https://arxiv.org/abs/2506.09820" target="_blank" rel="noopener">CoRT</a>ï¼æ code execution æ¾è¿ reasoningï¼æåä¸é´è®¡ç®çå¯éªè¯æ§ã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>é¡¹ç® / 代ç / benchmark</span><strong>Executable artifact stack</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="/projects/agentic-creation-executable-artifacts/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>Agentic Creation</strong><span>坿§è¡äº§ç©çæãéªè¯å交äºå¼è¯æµæ»è§ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/gamecraft-bench" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/gamecraft-bench</strong><span>å¯ç©æ¸¸æçæ benchmarkãä»»å¡ãreplay å hidden rubricã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/BlenderLLM" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/BlenderLLM</strong><span>CAD script çææ¨¡åãBlendNet æ°æ®å CADBenchã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/Mamo" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/Mamo</strong><span>æ°å¦å»ºæ¨¡ benchmark ä¸ solver èµæºã</span></span></a> <a class="project-stack-row" href="https://github.com/Cardinal-Operations/ORLM" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>Cardinal-Operations/ORLM</strong><span>èªå¨åä¼å建模è®ç»æ¡æ¶å IndustryOR benchmarkã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-education" id="ai-education" data-project-tags="AI Education;Reasoning & Agents;Healthcare AI" data-project-url="/projects/ai-education/"> <div class="project-program-heading"> <div class="project-eyebrow">Education agents, tutors, and learning simulation</div> <h2>AI Education</h2> <p> è¿æ¡çº¿æ LLM agent ä»âçé¢å·¥å
·âæ¨è¿å°å¯æå¦ãå¯éªç»ãå¯è¯æµçæè²æºè½ä½ï¼å
æ¬èªæè°èå¦ä¹ 婿ãLLM user simulatorãPlatoLM / SocraticChat çèæ ¼æåºå¼å¤è½®å¯¹è¯è®ç»ï¼ä»¥å EasyMED / SPBench ä¸ç AI æ ååç
人ã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/ai-education/">Open AI Education</a> <a class="project-action" href="https://github.com/FreedomIntelligence/PlatoLM" target="_blank" rel="noopener">Open PlatoLM</a> <a class="project-action" href="https://github.com/FreedomIntelligence/EasyMED" target="_blank" rel="noopener">Open EasyMED</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/media-roundup-2026/socratic-pipeline.png" alt="Socratic dialogue and education agent pipeline"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2506.09968" target="_blank" rel="noopener">SRLAgent</a>ï¼ç¨æ¸¸æåæºå¶å LLM assistance å¢å¼º self-regulated learningã</li> <li> <a href="https://arxiv.org/abs/2308.11534" target="_blank" rel="noopener">Large Language Model as a User Simulator</a>ï¼æ LLM ä½ä¸ºå¯æ§ç¨æ·æ¨¡æå¨ï¼ç¨äºå¤è½®å¯¹è¯è®ç»åè¯æµã</li> <li> <a href="https://openreview.net/forum?id=9nddtu94uX" target="_blank" rel="noopener">PlatoLM</a>ï¼éè¿èæ ¼æåºå¼è¿½é®ï¼æå¤è½®æå¦å¯¹è¯è½åè¸é¦è¿ LLMã</li> <li> <a href="https://arxiv.org/abs/2511.14783" target="_blank" rel="noopener">Human or LLM as Standardized Patients?</a>ï¼ç¨ EasyMED å SPBench è¯æµ AI æ ååç
人çå»å¦æè²ææã</li> <li> <a href="https://arxiv.org/abs/2602.05856" target="_blank" rel="noopener">It Talks Like a Patient, But Feels Different</a>ï¼ä» CHI / co-design è§è§ç ç©¶ AI æ ååç
人çå¦ä¹ ä½éªåè®¾è®¡éæ±ã
1</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>é¡¹ç® / æ°æ® / 代ç èµæº</span><strong>Education agent stack</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="/projects/ai-education/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>AI Education</strong><span>æè²æºè½ä½ãå¦ä¹ 模æãSocratic dialogue å AI standardized patients æ»è§ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/PlatoLM" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/PlatoLM</strong><span>PlatoLM 模åãSocratic-style dialogue training å弿ºèµæºã</span></span></a> <a class="project-stack-row" href="https://huggingface.co/datasets/FreedomIntelligence/SocraticChat" target="_blank" rel="noopener"><span class="project-stack-type">HF Dataset</span><span class="project-stack-main"><strong>SocraticChat</strong><span>èæ ¼æåºå¼å¯¹è¯æ°æ®èµæºã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/EasyMED" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/EasyMED</strong><span>AI æ ååç
人åå»å¦æè²è¯æµèµæºã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-economics" id="economic-world-models" data-project-tags="World Models & Simulation;Reasoning & Agents" data-project-url="/projects/economic-world-models/"> <div class="project-program-heading"> <div class="project-eyebrow">Economic world models, financial LLMs, and agent applications</div> <h2>ç»æµä¸ç模ååéè大模ååå
¶ Agent åºç¨</h2> <p> è¿æ¡ç 究主线æå¸åºãæºæãæ¿çå夿ºè½ä½è¡ä¸ºå»ºæå¯è®¡ç®ã坿¼åãå¯å¯¹é½çç»æµä¸çï¼åæ¶è¦çéè大模åãéèè½åè¯æµã夿¨¡æéèæ¨¡åãä¸è±åè¯éèæ¨¡åï¼ä»¥åé¢åçå®å¸åºç Agent åºç¨ã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/economic-world-models/">Open Economic World Models</a> <a class="project-action" href="https://github.com/FreedomIntelligence/Awesome-Econ-World-Models/" target="_blank" rel="noopener">Awesome EconWM</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/economic-world-models/ewm-hero.png" alt="Economic World Models compare physical-world transitions with agent-generated economic-world transitions"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="/assets/pdf/economic-world-models.pdf" target="_blank" rel="noopener">From Economic Agents to Agentic Economies</a>ï¼æåºç»æµä¸ç模åçç³»ç»èå¾ï¼ç¨å¯è®¡ç®ç»æµä½æ¯ææºè½ä½è®ç»ãæ¿çæ²çãè§ååå®å
¨åæã</li> <li> <a href="https://arxiv.org/abs/2502.01506" target="_blank" rel="noopener">TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets</a>ï¼ç¨ LLM æèµè
模æéèå¸åºä¸ç信念ãæå¾ãä¿¡æ¯æµå交æè¡ä¸ºã</li> <li> <a href="https://aclanthology.org/2025.findings-naacl.300/" target="_blank" rel="noopener">UCFE: A User-Centric Financial Expertise Benchmark for Large Language Models</a>ï¼ä»ç¨æ·éæ±å卿éèä»»å¡åºåè¯æµå¤§æ¨¡åçéèä¸ä¸è½åã</li> <li> <a href="https://arxiv.org/abs/2402.12659" target="_blank" rel="noopener">FinBen: A Holistic Financial Benchmark for Large Language Models</a>ï¼é¢åå¤ç±»éèä»»å¡ç综å大è¯è¨æ¨¡åè¯æµåºåã</li> <li> <a href="https://arxiv.org/abs/2408.11878" target="_blank" rel="noopener">Open-FinLLMs: Open Multimodal Large Language Models for Financial Applications</a>ï¼å¼æ¾é¢åéèåºç¨ç夿¨¡æå¤§æ¨¡åãæ°æ®åè¯æµä½ç³»ã</li> <li> <a href="https://arxiv.org/abs/2403.06249" target="_blank" rel="noopener">No Language Is an Island</a>ï¼ç»ä¸ä¸è±æéè大模åãæä»¤æ°æ®ä¸ benchmarkï¼æ¨å¨åè¯éèæºè½ã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>GitHub / Project èµæº</span><strong>Paper, repositories, project pages</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="/projects/economic-world-models/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>Economic World Models</strong><span>æ»è§ç»æµä¸çæ¨¡åæ¶æãè½åé¶æ¢¯ã代表å¢éåç³»ç»è·¯çº¿å¾ã</span></span></a> <a class="project-stack-row" href="/assets/pdf/economic-world-models.pdf" target="_blank" rel="noopener"><span class="project-stack-type">Paper</span><span class="project-stack-main"><strong>From Economic Agents to Agentic Economies</strong><span>ç»æµä¸ç模å论æ PDFã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/Awesome-Econ-World-Models/" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/Awesome-Econ-World-Models</strong><span>ç»æµä¸ç模å论æãç³»ç»åèµæºéåã</span></span></a> <a class="project-stack-row" href="/projects/twinmarket-financial-simulation/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>TwinMarket Financial AI Simulation</strong><span>ç»æµä¸ç模åä¸çéèå¸åºååæ²çã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/TwinMarket" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/TwinMarket</strong><span>éèè¡ä¸ºä¸ç¤¾ä¼æ¨¡æä»£ç èµæºã</span></span></a> <a class="project-stack-row" href="/projects/microverse-world-simulation/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>MicroVerse Micro-World Simulation</strong><span>æä¸çæ¨¡åæ©å±å°ç§å¦å¾®ä¸çä¸çç©è¿ç¨ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/MicroVerse" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/MicroVerse</strong><span>MicroVerse 项ç®èµæºã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-optimization" id="optimization-llms" data-project-tags="Reasoning & Agents;AI Systems &
1 Infrastructure" data-project-url="/projects/reasoning-math-optimization/"> <div class="project-program-heading"> <div class="project-eyebrow">Optimization modeling LLMs</div> <h2>è¿ç¹ä¼å大模åç³»å</h2> <p> è¿ç¹ä¼å主线æèªç¶è¯è¨é®é¢è½¬ææ°å¦ä¼å模åã约æãç®æ 彿°å坿§è¡ä»£ç ãORLM æä¾å¼æ¾æ¨¡åä¸ IndustryOR benchmarkï¼CALM/STORM è¿ä¸æ¥ç ç©¶ä¼å建模ä¸çåçæ¨çè½åã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/reasoning-math-optimization/">Open Reasoning and Optimization</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/freedomai-2026/impact/iclr-financial-ai-award.jpg" alt="Reasoning and optimization project signal"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2405.17743" target="_blank" rel="noopener">ORLM: A Customizable Framework in Training Large Models for Automated Optimization Modeling</a>ï¼OR-InstructãORLM å IndustryORã</li> <li> <a href="https://arxiv.org/abs/2510.04204" target="_blank" rel="noopener">CALM Before the STORM: Unlocking Native Reasoning for Optimization Modeling</a>ï¼ä¼å建模ä¸çåçæ¨çã</li> <li> <a href="https://arxiv.org/abs/2403.02884" target="_blank" rel="noopener">MathScale: Scaling Instruction Tuning for Mathematical Reasoning</a>ï¼æ°å¦æ¨ç scaling ä¸ verifier ç¸å
³åºç¡ã</li> <li> <a href="https://arxiv.org/abs/2506.09820" target="_blank" rel="noopener">CoRT: Code-integrated Reasoning within Thinking</a>ï¼å°ä»£ç æ§è¡æ¥å
¥æç»´é¾ï¼ä¹æå¡äºå½¢å¼å建模å计ç®ã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>GitHub / Hugging Face èµæº</span><strong>Code, models, datasets</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="https://github.com/Cardinal-Operations/ORLM" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>Cardinal-Operations/ORLM</strong><span>èªå¨åä¼å建模è®ç»æ¡æ¶ãæ°æ®åæå benchmark å
¥å£ã</span></span></a> <a class="project-stack-row" href="https://huggingface.co/datasets/CardinalOperations/IndustryOR" target="_blank" rel="noopener"><span class="project-stack-type">HF Dataset</span><span class="project-stack-main"><strong>CardinalOperations/IndustryOR</strong><span>100 个çå®å·¥ä¸è¿ç¹é®é¢ benchmarkã</span></span></a> <a class="project-stack-row" href="https://github.com/tangzhy/STORM" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>tangzhy/STORM</strong><span>CALM/STORM ä¼å建模æ¨çèµæºã</span></span></a> <a class="project-stack-row" href="https://github.com/ChengpengLi1003/CoRT" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>ChengpengLi1003/CoRT</strong><span>代ç éææ¨çèµæºã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-optimization" id="legal-ai" data-project-tags="Legal AI;Reasoning & Agents;Evaluation & Interpretability" data-project-url="/projects/legal-ai-models-agents/"> <div class="project-program-heading"> <div class="project-eyebrow">Foundation models, user-centric evaluation, evidence retrieval, and legal agents</div> <h2> <span data-lang="en">Legal AI Models, Retrieval, and Agents</span><span data-lang="zh">æ³å¾å¤§æ¨¡åãæ£ç´¢ä¸æºè½ä½</span> </h2> <p><span data-lang="en">A research path from HanFei-1.0 and UCL-Bench to GenArtID, Legal-R1, and continuously improving Legal AI systems connected to real lawyer workflows.</span><span data-lang="zh">ä» HanFei-1.0 ä¸ UCL-Benchï¼å° GenArtIDãLegal-R1 åé¢åçå®å¾å¸å·¥ä½æµçæ³å¾ AI éç¯ã</span></p> <div class="project-actions"> <a class="project-action primary" href="/projects/legal-ai-models-agents/">Open Legal AI Project</a> <a class="project-action" href="https://aclanthology.org/2025.findings-naacl.444/" target="_blank" rel="noopener">UCL-Bench</a> <a class="project-action" href="https://openreview.net/forum?id=2Oe3FRRcVQ" target="_blank" rel="noopener">Legal-R1</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"><img src="/assets/img/legal-ai/joint-lab-signing.png" alt="CUHK-Shenzhen and FaDaFu Legal AI Joint Laboratory signing ceremony"></div> <div> <h3> <span data-lang="en">One connected research program</span><span data-lang="zh">䏿¡è¿è´¯çææ¯è·¯çº¿</span> </h3> <ol class="project-paper-list"> <li> <strong>HanFei-1.0:</strong> Chinese legal foundation-model training.</li> <li> <strong>UCL-Bench:</strong> 5 legal scenarios and 22 user-centric tasks.</li> <li> <strong>GenArtID:</strong> generative statutory article identifier prediction.</li> <li> <strong>Legal-R1:</strong> multi-step agentic retrieval for evidence-based reasoning.</li> <li> <strong>Legal AI within a Loop:</strong> Harness â Benchmark â Training â Data Flywheel.</li> </ol> </div> </div> </section> <section class="project-program-section project-theme-medical" id="huatuogpt" data-project-tags="Healthcare AI;Foundation Models" data-project-url="/projects/medical-ai-ecosystem/"> <div class="project-program-heading"> <div class="project-eyebrow">HuatuoGPT medical model family</div> <h2>AI for Healthcareï¼åä½GPTç³»å大模å</h2> <p> åä½GPT æ¯ AI for Healthcare 主线ï¼ä» Huatuo-26M 䏿å»å¦é®çæ°æ®ï¼å° HuatuoGPT / HuatuoGPT-II å»å¦éé
ï¼åå° HuatuoGPT-o1 ç夿å»å¦æ¨çå HuatuoGPT-Vision çå»å¦è§è§è¯è¨çè§£ã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/medical-ai-ecosystem/">Open HuatuoGPT Ecosystem</a> <a class="project-action" href="/projects/huatuogpt-ii/">Open HuatuoGPT-II</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/media-roundup-2026/huatuogpt-overview.png" alt="HuatuoGPT medical model ecosystem"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2305.01526" target="_blank" rel="noopener">Huatuo-26M, a Large-scale Chinese Medical QA Dataset</a>ï¼ä¸æå»å¦ QA æ°æ®åºç¡ã</li> <li> <a href="https://arxiv.org/abs/2305.15075" target="_blank" rel="noopener">HuatuoGPT, towards Taming Language Model to Be a Doctor</a>ï¼æ©æå¼æ¾å»å¦å¯¹è¯æ¨¡åã</li> <li>
1 <a href="https://arxiv.org/abs/2311.09774" target="_blank" rel="noopener">HuatuoGPT-II, One-stage Training for Medical Adaption of LLMs</a>ï¼ç»ä¸ä¸é¶æ®µå»å¦éé
è®ç»ã</li> <li> <a href="https://arxiv.org/abs/2412.18925" target="_blank" rel="noopener">Towards Medical Complex Reasoning with LLMs through Medical Verifiable Problems</a>ï¼HuatuoGPT-o1 夿å»å¦æ¨çã</li> <li> <a href="https://arxiv.org/abs/2406.19280" target="_blank" rel="noopener">HuatuoGPT-Vision, Towards Injecting Medical Visual Knowledge into Multimodal LLMs at Scale</a>ï¼å»å¦å¤æ¨¡æè§è§ç¥è¯æ³¨å
¥ã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>GitHub / Hugging Face èµæº</span><strong>Code, models, datasets</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/Huatuo-26M" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/Huatuo-26M</strong><span>䏿å»å¦ QA æ°æ®éä¸æå»ºèµæºã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/HuatuoGPT" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/HuatuoGPT</strong><span>åä½GPT åºç¡å»å¦å¯¹è¯æ¨¡å代ç ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/HuatuoGPT-II" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/HuatuoGPT-II</strong><span>HuatuoGPT-II è®ç»åè¯æµææã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/HuatuoGPT-o1" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/HuatuoGPT-o1</strong><span>å»å¦å¤ææ¨ç模ååæ°æ®ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/HuatuoGPT-Vision" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/HuatuoGPT-Vision</strong><span>å»å¦è§è§è¯è¨æ¨¡åè®ç»ä»£ç ã</span></span></a> <a class="project-stack-row" href="https://huggingface.co/datasets/FreedomIntelligence/PubMedVision" target="_blank" rel="noopener"><span class="project-stack-type">HF Dataset</span><span class="project-stack-main"><strong>FreedomIntelligence/PubMedVision</strong><span>å»å¦å¾æè§è§æä»¤æ°æ®ã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-eval" id="medical-evaluation" data-project-tags="Healthcare AI;Evaluation & Interpretability" data-project-url="/projects/medical-evaluation-benchmarks/"> <div class="project-program-heading"> <div class="project-eyebrow">Healthcare AI evaluation</div> <h2>å»ç AI è¯æµç³»å</h2> <p> å»ç AI ä¸è½åªçè天æµç
度ãè¿æ¡çº¿æè¯æµææä¸æå»å¦ç¥è¯ã夿¨¡æå»å¦çè§£ãå¯è§£éè¯æã宿¶ä¸´åºãå»ç工使µåå»å¦æè²å å±ï¼è®©æ¨¡å卿´æ¥è¿ä¸´åºçä»»å¡ä¸æ´é²çæ¿ã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/medical-evaluation-benchmarks/">Open Medical Evaluation Benchmarks</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/media-roundup-2026/cmb-overview.png" alt="CMB medical benchmark overview"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2308.08833" target="_blank" rel="noopener">CMB: A Comprehensive Medical Benchmark in Chinese</a>ï¼ä¸æå»å¦ç»¼å benchmarkã</li> <li> <a href="https://arxiv.org/abs/2408.03361" target="_blank" rel="noopener">GMAI-MMBench</a>ï¼é¢åéç¨å»ç AI ç夿¨¡æè¯æµã</li> <li> <a href="https://arxiv.org/abs/2407.13301" target="_blank" rel="noopener">Chain-of-Diagnosis / DxBench</a>ï¼å¯è§£éè¯ææ¨¡ååçå®å»æ£å¯¹è¯è¯æè¯æµã</li> <li> <a href="https://arxiv.org/abs/2602.16747" target="_blank" rel="noopener">LiveClin: A Live Clinical Benchmark without Leakage</a>ï¼æç»æ´æ°ãææ³æ¼ççå®ä¸´åº benchmarkã</li> <li> <a href="https://arxiv.org/abs/2510.11040" target="_blank" rel="noopener">Enabling Doctor-Centric Medical AI with LLMs through Workflow-Aligned Tasks and Benchmarks</a>ï¼DoctorFLAN / DotaBenchï¼é¢åå»ç工使µã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>GitHub / Hugging Face èµæº</span><strong>Code, models, datasets</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/CMB" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/CMB</strong><span>
1䏿å»å¦ benchmark 代ç ã</span></span></a> <a class="project-stack-row" href="https://github.com/uni-medical/GMAI-MMBench" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>uni-medical/GMAI-MMBench</strong><span>GMAI-MMBench 夿¨¡æè¯æµä»£ç ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/Chain-of-Diagnosis" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/Chain-of-Diagnosis</strong><span>CoDãDxBenchãDiagnosisGPT ç»ä¸å
¥å£ã</span></span></a> <a class="project-stack-row" href="https://github.com/AQ-MedAI/LiveClin" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>AQ-MedAI/LiveClin</strong><span>LiveClin 代ç åæ°æ®å
¥å£ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/DotaGPT" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/DotaGPT</strong><span>DoctorFLAN / DotaBench è®ç»ä¸è¯æµä»£ç ã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-education" id="medical-education-sp" data-project-tags="Healthcare AI;AI Education;Reasoning & Agents" data-project-url="/projects/medical-education-standardized-patients/"> <div class="project-program-heading"> <div class="project-eyebrow">Medical education and AI standardized patients</div> <h2>å»å¦æè²ä¸ AI æ ååç
人</h2> <p> è¿æ¡çº¿é¢åå»å¦æè²ä¸ç SPï¼standardized patientsï¼è®ç»ï¼æ AI ç
人åæå¯æ§ãå¯å¤ç¨ãå¯è¯æµçç»ä¹ 对象ãå®åæ¶è¿æ¥ ACL å»å¦æè²è¯æµä¸ CHI/人æºäº¤äºéçå
±å设计é®é¢ï¼AI ç
人ä¸åªè¦âåç
人说è¯âï¼è¿è¦çæ£å¸®å©å¦çç»ä¹ ç
å²ééãæ²éåä¸´åºæ¨çã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/medical-education-standardized-patients/">Open AI Standardized Patients</a> <a class="project-action" href="https://github.com/FreedomIntelligence/EasyMED" target="_blank" rel="noopener">Open EasyMED</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/media-roundup-2026/huatuogpt-auto-eval.jpeg" alt="Medical education and standardized patient evaluation"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2511.14783" target="_blank" rel="noopener">Human or LLM as Standardized Patients? A Comparative Study in Medical Education</a>ï¼æåº EasyMED ä¸ SPBenchï¼å¯¹æ¯ AI SP ä¸äººç±» SP çæå¦ææã</li> <li> <a href="https://arxiv.org/abs/2602.05856" target="_blank" rel="noopener">"It Talks Like a Patient, But Feels Different": Co-Designing AI Standardized Patients with Medical Learners</a>ï¼ä» CHI è§è§æ»ç»å»å¦å¦ä¹ è
对 AI SP çä½éªåè®¾è®¡éæ±ã</li> <li> <a href="https://arxiv.org/abs/2510.11040" target="_blank" rel="noopener">Doctor-centric workflow-aligned tasks and benchmarks</a>ï¼æå»å¦ AI è¯æµæ¾è¿å»ççå®å·¥ä½æµã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>GitHub / 论æèµæº</span><strong>Code and papers</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/EasyMED" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/EasyMED</strong><span>AI æ ååç
äººæ¡æ¶åå»å¦æè²è¯æµå
¥å£ã</span></span></a> <a class="project-stack-row" href="https://arxiv.org/abs/2511.14783" target="_blank" rel="noopener"><span class="project-stack-type">Paper</span><span class="project-stack-main"><strong>Human or LLM as Standardized Patients?</strong><span>EasyMEDãSPBench åå»å¦æè²å¯¹æ¯å®éªã</span></span></a> <a class="project-stack-row" href="https://dl.acm.org/doi/10.1145/3772363.3798336" target="_blank" rel="noopener"><span class="project-stack-type">CHI</span><span class="project-stack-main"><strong>Co-designing AI standardized patients</strong><span>
1å»å¦å¦ä¹ è
å
±å设计ä¸äº¤äºéæ±ã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-multilingual" id="multilingual-llms" data-project-tags="Foundation Models;Healthcare AI" data-project-url="/projects/multilingual-localized-llms/"> <div class="project-program-heading"> <div class="project-eyebrow">Multilingual and localized LLMs</div> <h2>å¤è¯è¨å¤§è¯è¨æ¨¡å</h2> <p> å¤è¯è¨ä¸»çº¿ä» Phoenix / LLMZoo ç弿¾å¤è¯è¨è天å¼å§ï¼å° AceGPT çé¿æä¼¯è¯æ¬å°åï¼åå° Apollo / ApolloMoE çå¤è¯è¨å»å¦æ¨¡åãè¯æåè¯æµãæ ¸å¿ç®æ æ¯è®©å»å¦åéç¨ AI ä¸åªæå¡å°æ°é«èµæºè¯è¨ã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/multilingual-localized-llms/">Open Multilingual Project</a> <a class="project-action" href="/projects/rag-instruction-data/">Open Data Infrastructure</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/freedomai-2026/impact/apollo-map.jpg" alt="Apollo multilingual medical model map"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2304.10453" target="_blank" rel="noopener">Phoenix: Democratizing ChatGPT across Languages</a>ï¼æ©æå¼æ¾å¤è¯è¨è天模åã</li> <li> <a href="https://aclanthology.org/2024.naacl-long.450/" target="_blank" rel="noopener">AceGPT, Localizing Large Language Models in Arabic</a>ï¼é¿æä¼¯è¯å¤§æ¨¡åæ¬å°åä¸å¯¹é½ã</li> <li> <a href="https://arxiv.org/abs/2403.03640" target="_blank" rel="noopener">Apollo: Lightweight Multilingual Medical LLMs towards Democratizing Medical AI to 6B People</a>ï¼å¤è¯è¨å»å¦æ¨¡åãæ°æ®åè¯æµã</li> <li> <a href="https://arxiv.org/abs/2409.02615" target="_blank" rel="noopener">ApolloMoE</a>ï¼é¢å 50 ç§è¯è¨çå»å¦ MoE 模åã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>GitHub / Hugging Face èµæº</span><strong>Code, models, datasets</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/LLMZoo" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/LLMZoo</strong><span>Phoenix / LLMZoo å¤è¯è¨æ¨¡åèµæºã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/AceGPT" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/AceGPT</strong><span>é¿æä¼¯è¯ LLM æ¬å°å代ç åèµæºã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/Apollo" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/Apollo</strong><span>Apollo å¤è¯è¨å»å¦æ¨¡åãæ°æ®ãè¯æµä»£ç ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/ApolloMoE" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/ApolloMoE</strong><span>é¢å 50 ç§è¯è¨çå»å¦ MoE 模å代ç ã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-speech" id="speech-llms" data-project-tags="Foundation Models;Reasoning & Agents" data-project-url="/projects/speech-audio-talking-head-ai/"> <div class="project-program-heading"> <div class="project-eyebrow">Speech LLMs and speech-to-speech agents</div> <h2>Speech LLMs</h2> <p> Speech LLMs æ¹åæå¤§æ¨¡åä»ææ¬æ©å±å°è¯é³è¾å
¥ãè¯é³è¾åºãè¯æ°ãæ
绪åå¯è¯è¨ä¿¡æ¯ãç¸å
³é¡¹ç®å
æ¬ SoundwaveãS2S-ArenaãEchoMindãFusionAudioãUNSPOKENï¼ä»¥å speech-to-speech Turing Testã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/speech-audio-talking-head-ai/">Open Speech and Audio Project</a> <a class="project-action" href="https://github.com/FreedomIntelligence/Soundwave" target="_blank" rel="noopener">Open Soundwave</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/media-roundup-2026/socratic-pipeline.png" alt="Speech LLM and interactive AI pipeline"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2502.12900" target="_blank" rel="noopener">Soundwave: Less is More for Speech-Text Alignment in LLMs</a>ï¼ç¨æ´å°æ°æ®å®æ speech-text alignmentã</li> <li> <a href="https://arxiv.org/abs/2503.05085" target="_blank" rel="noopener">S2S-Arena</a>ï¼è¯æµ speech-to-speech 模åçè¯ä¹çè§£åå¯è¯è¨æä»¤è·éã</li> <li> <a href="https://arxiv.org/abs/2510.22758" target="_blank" rel="noopener">EchoMind</a>ï¼é¢å empathetic Speech Language Models çå¤å±çº§ benchmarkã</li> <li> <a href="https://arxiv.org/abs/2602.24080" target="_blank" rel="noopener">Human or Machine? A Preliminary Turing Test for Speech-to-Speech Interaction</a>ï¼è¯æµ S2S ç³»ç»æ¯å¦çæ£å人类对è¯ã</li> <li> <a href="https://arxiv.org/abs/2506.01111" target="_blank" rel="noopener">FusionAudio-1.2M</a>ï¼ç»ç²åº¦é³é¢ caption ä¸ multimodal contextual fusionã</li> <li> <a href="https://github.com/Hongru0306/UNSPOKEN" target="_blank" rel="noopener">UNSPOKEN</a>ï¼ç¨æªè¯´åºå£ç声å¦çº¿ç´¢è¯æµ audio language models çé廿¨çã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>GitHub / Hugging Face èµæº</span><strong>Code, models, datasets</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/Soundwave" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/Soundwave</strong><span>Speech-text alignment for LLMsã</span></span></a> <a class="project-stack-row" href="https://huggingface.co/FreedomIntelligence/Soundwave" target="_blank" rel="noopener"><span class="project-stack-type">HF Model</span><span class="project-stack-main"><strong>FreedomIntelligence/Soundwave</strong><span>Soundwave 模åå¡ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/S2S-Arena" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/S2S-Arena</strong><span>Speech-to-speech paralinguistic benchmarkã</span></span></a> <a class="project-stack-row" href="https://hlt-cuhksz.github.io/EchoMind/" target="_blank" rel="noopener"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>EchoMind</strong><span>Empathetic speech language model benchmarkã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/FusionAudio" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/FusionAudio</strong><span>Fine-grained audio captioning data and modelsã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-explain" id="interpretable-llms" data-project-tags="Evaluation & Interpretability;Healthcare AI" data-project-url="/projects/llm-interpretability/"> <div class="project-program-heading"> <div class="project-eyebrow">Explainable and verifiable LLMs</div> <h2>LLM Interpretability</h2> <p> å¯è§£éæ§æ¯æ¨ªåè½åï¼å®å
³å¿æ¨¡å为ä»ä¹è¿æ ·åçãåªäºå
é¨ç¹å¾çæ£æç¨ã以åå¦ä½ææºå¶ç解转æå¯æ§ç¼è¾ãè¿ä¸ªæ¹åç°å¨ä»¥ sparse autoencoderãmodel unlearning å circuit analysis 为主线ï¼åæ¶è¿æ¥å»å¦è¯æé¾ãverifier å代ç éææ¨çã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/llm-interpretability/">Open LLM Interpretability</a> <a class="project-action" href="https://arxiv.org/abs/2510.03659" target="_blank" rel="noopener">Open ICLR 2026 paper</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/freedomai-2026/impact/llm-interpretability.svg" alt="LLM interpretability and sparse autoencoder analysis"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2510.03659" target="_blank" rel="noopener">Does Higher Interpretability Imply Better Utility? A Pairwise Analysis on Sparse Autoencoders</a>ï¼ICLR 2026ï¼ç³»ç»æ¯è¾ SAE çå¯è§£éæ§ææ åå®é
steering utilityï¼è· NeurIPS 2025 ResponsibleFM Workshop Outstanding Paper Awardã
1</li> <li> <a href="https://arxiv.org/abs/2505.24428" target="_blank" rel="noopener">Model Unlearning via Sparse Autoencoder Subspace Guided Projections</a>ï¼EMNLP 2025ï¼ç¨ SAE subspace guided projections 忴坿§ç模åéå¿ã</li> <li> <a href="https://arxiv.org/abs/2502.11812" target="_blank" rel="noopener">Towards Understanding Fine-Tuning Mechanisms of LLMs via Circuit Analysis</a>ï¼ICML 2025ï¼ç¨ circuit analysis ç解微è°å¦ä½æ¹å模åå
é¨è®¡ç®ã</li> <li> <a href="https://arxiv.org/abs/2407.13301" target="_blank" rel="noopener">CoD, Towards an Interpretable Medical Agent using Chain of Diagnosis</a>ï¼ç¨äºæ¥è¯æé¾å置信度å叿åå»å¦è¯æå¯è§£éæ§ã</li> <li> <a href="https://arxiv.org/abs/2311.09724" target="_blank" rel="noopener">Outcome-supervised Verifiers for Planning in Mathematical Reasoning</a>ï¼ç¨ç»æçç£ verifier è¾
婿°å¦è§åæ¨çã</li> <li> <a href="https://arxiv.org/abs/2506.09820" target="_blank" rel="noopener">CoRT: Code-integrated Reasoning within Thinking</a>ï¼æä»£ç æ§è¡æ¾å
¥æç»´è¿ç¨ï¼è®©ä¸é´è®¡ç®æ´éæã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>论æ / GitHub / Hugging Face èµæº</span><strong>Papers, code, models, datasets</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="/projects/llm-interpretability/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>LLM Interpretability and Sparse Autoencoders</strong><span>SAE utilityãSAE-guided unlearning å circuit analysis ç详æ
页ã</span></span></a> <a class="project-stack-row" href="https://arxiv.org/abs/2510.03659" target="_blank" rel="noopener"><span class="project-stack-type">ICLR 2026</span><span class="project-stack-main"><strong>Does Higher Interpretability Imply Better Utility?</strong><span>Pairwise analysis on sparse autoencodersã</span></span></a> <a class="project-stack-row" href="https://arxiv.org/abs/2505.24428" target="_blank" rel="noopener"><span class="project-stack-type">EMNLP 2025</span><span class="project-stack-main"><strong>Model Unlearning via SAE Subspace Guided Projections</strong><span>å¯è§£éå空é´å¼å¯¼ç模åéå¿ã</span></span></a> <a class="project-stack-row" href="https://arxiv.org/abs/2502.11812" target="_blank" rel="noopener"><span class="project-stack-type">ICML 2025</span><span class="project-stack-main"><strong>Fine-Tuning Mechanisms via Circuit Analysis</strong><span>ç¨ circuit analysis çè§£ LLM å¾®è°æºå¶ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/Chain-of-Diagnosis" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/Chain-of-Diagnosis</strong><span>CoD 代ç ãè¯æé¾æ°æ®ãç¾ç
åºãDxBench å DiagnosisGPT å
¥å£ã</span></span></a> <a class="project-stack-row" href="https://huggingface.co/FreedomIntelligence/DiagnosisGPT-6B" target="_blank" rel="noopener"><span class="project-stack-type">HF Model</span><span class="project-stack-main"><strong>FreedomIntelligence/DiagnosisGPT-6B</strong><span>åºäº CoD çå¯è§£éè¯ææ¨¡åã</span></span></a> <a class="project-stack-row" href="https://huggingface.co/datasets/FreedomIntelligence/DxBench" target="_blank" rel="noopener"><span class="project-stack-type">HF Dataset</span><span class="project-stack-main"><strong>FreedomIntelligence/DxBench</strong><span>çå®å»æ£å¯¹è¯æå»ºçç¾ç
è¯æ benchmarkã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/OVM" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/OVM</strong><span>Outcome-supervised verifier ç¸å
³ä»£ç ã</span></span></a> <a class="project-stack-row" href="https://github.com/ChengpengLi1003/CoRT" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>ChengpengLi1003/CoRT</strong><span>代ç éææ¨ç项ç®ã</span></span></a> <a class="project-stack-row" href="https://github.com/tulerfeng/Video-R1" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>tulerfeng/Video-R1</strong><span>è§é¢æ¨ç强åå¦ä¹ 代ç ã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-eval" id="general-llm-evaluation" data-project-tags="Evaluation & Interpretability;Foundation Models;Healthcare AI" data-project-url="/projects/general-llm-evaluation/"> <div class="project-program-heading"> <div class="project-eyebrow">Rubrics, judges, freshness, bias, ethics, and rewards</div> <h2>éç¨å¤§æ¨¡åè¯ä¼°</h2> <p> è¿ä¸ªæ¹åæ Rubric-based Evaluation & Reward Design å Trustworthy / Responsible AI Evaluation åå¹¶æä¸æ¡æ´å¤§çâéç¨è¯ä¼°â主线ï¼ä» per-sample criteriaãLLM-as-a-judge åå·®ãæ¶é´æ³åãæååå·®ï¼å°å»ç伦çå rubric-to-rewardï¼æå¼æ¾å¼è½åè¯æµåæå¯å®¡è®¡ãå¯å¤ç°ãå¯è®ç»çåé¦ç³»ç»ã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/general-llm-evaluation/">Open General Evaluation</a> <a class="project-action" href="https://github.com/FreedomIntelligence/Awesome-Rubrics" target="_blank" rel="noopener">Open Awesome-Rubrics</a> <a class="project-action" href="https://github.com/FreedomIntelligence/MLLM-Bench" target="_blank" rel="noopener">Open MLLM-Bench</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/freedomai-2026/impact/github-impact.jpg" alt="General LLM evaluation project signal"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://aclanthology.org/2025.naacl-long.256/" target="_blank" rel="noopener">MLLM-Bench</a>ï¼ç¨ per-sample criteria è¯ä¼°å¤æ¨¡æå¤§æ¨¡åï¼è®©æ ·æ¬çº§æ åæä¸º judge çæ ¸å¿è¾å
¥ã</li> <li> <a href="https://arxiv.org/abs/2402.10669" target="_blank" rel="noopener">Humans or LLMs as the Judge?</a>ï¼ç³»ç»åæäººç±»å LLM judge ç夿åå·®ï¼ä¸ºèªå¨è¯æµæ ¡åæä¾ä¾æ®ã</li> <li> <a href="https://github.com/FreedomIntelligence/FreshBench" target="_blank" rel="noopener">FreshBench</a>ï¼è¯ä¼°å¤§æ¨¡åç temporal generalizationï¼å¤ææ¨¡åæ¯å¦âè¿æ¶âã</li> <li> <a href="https://arxiv.org/abs/2505.18562" target="_blank" rel="noopener">From Word to World</a>ï¼ç¨ word association test è¯ä¼°åç¼è§£ LLM çæååå·®ã</li> <li> <a href="https://github.com/FreedomIntelligence/PrinciplismQA" target="_blank" rel="noopener">PrinciplismQA</a>ï¼æåååå»å¦ä¼¦ç转æ MCQAã临åºç
ä¾å rubric keypointsã</li> <li> <a href="https://github.com/FreedomIntelligence/Awesome-Rubrics" target="_blank" rel="noopener">Awesome-Rubrics</a>ï¼æ´ç rubric-based evaluationãreward modeling å post-training èµæºã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>论æ / GitHub / èµæºå
¥å£</span><strong>Evaluation and reward stack</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="/projects/general-llm-evaluation/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>General LLM Evaluation</strong><span>éç¨å¤§æ¨¡åè¯ä¼°ãrubric judgeãå¯é æ§å reward design æ»è§ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/Awesome-Rubrics" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/Awesome-Rubrics</strong><span>Rubric-based evaluationãalignment å agentic AI æç®å°å¾ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/MLLM-Bench" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/MLLM-Bench</strong><span>Per-sample criteria 夿¨¡æè¯æµä»£ç 䏿°æ®ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/FreshBench" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/FreshBench</strong><span>
1æ¶é´æ³å忍¡åæ°é²åº¦è¯æµèµæºã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/PrinciplismQA" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/PrinciplismQA</strong><span>å»å¦ä¼¦çç¥è¯ãç
ä¾å rubric scoring èµæºã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-medical" id="clinical-agents-hospital-world-models" data-project-tags="Healthcare AI;World Models & Simulation;Reasoning & Agents" data-project-url="/projects/clinical-agents-hospital-world-models/"> <div class="project-program-heading"> <div class="project-eyebrow">Clinical agents, patient dynamics, HIS environments, and hospital twins</div> <h2>Clinical & Hospital World Models</h2> <p> è¿ä¸ªæ¹åæå»ç AI ä»éæé®çæ¨è¿å°ä¸´åºç¯å¢ä¸ç agentï¼å®å
³å¿ ICU æ£è
卿ãå»é¢ä¿¡æ¯ç³»ç»ãå»ç工使µãå¤è§è²å»é¢ä»¿çãå·¥å
·è°ç¨åå®å
¨æ²»çãçç§ benchmark å DentalGPT å·²åç¬æ¾å
¥çç§å¤§æ¨¡åæ¹åï¼è¿éåªä¿ç临åºä¸å»é¢ä¸ç模åã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/clinical-agents-hospital-world-models/">Open Clinical & Hospital Models</a> <a class="project-action" href="https://github.com/FreedomIntelligence/SepsisAgent" target="_blank" rel="noopener">Open SepsisAgent</a> <a class="project-action" href="https://github.com/FreedomIntelligence/TwinHospital" target="_blank" rel="noopener">Open TwinHospital</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/media-roundup-2026/huatuogpt-auto-eval.jpeg" alt="Clinical agents and hospital world models"> </div> <div> <h3>ç¸å
³è®ºææ´ç</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2605.14723" target="_blank" rel="noopener">SepsisAgent</a>ï¼éè¿ Clinical World Model è¿è¡ propose-simulate-refineï¼å¹¶ç¨ agentic RL å¦ä¹ ICU æ²»ççç¥ã</li> <li> <a href="https://github.com/FreedomIntelligence/Agentic-Hospital" target="_blank" rel="noopener">Agentic-Hospital</a>ï¼æ HIS/EMRãè§è²ä»¿çãå·¥å
·æ¥å£ãæ²»çå rubric è¯æµç»ç»æå»é¢çº§ agent ç¯å¢ã</li> <li> <a href="https://github.com/FreedomIntelligence/TwinHospital" target="_blank" rel="noopener">TwinHospital</a>ï¼ç»åçå® HIS data flow ä¸å¤è§è²ä»¿çï¼å¤ç°æå·ãé®è¯ãæ²»çåç®¡çæµç¨ã</li> <li> <a href="https://arxiv.org/abs/2510.11040" target="_blank" rel="noopener">Doctor-Centric Medical AI</a>ï¼æä»»å¡å benchmark 对é½å°å»ççå®å·¥ä½æµã</li> <li> <a href="https://github.com/FreedomIntelligence/OpenClaw-Medical-Skills" target="_blank" rel="noopener">OpenClaw-Medical-Skills</a>ï¼å¼æ¾å»ç AI skills libraryï¼ä¸ºä¸´åº agent æä¾å¯è°ç¨æè½å
¥å£ã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>ä¸´åº agent / å»é¢ç¯å¢ / æè½åº</span><strong>Hospital world-model stack</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="/projects/clinical-agents-hospital-world-models/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>Clinical & Hospital World Models</strong><span>临åºä¸å»é¢æ°åç¯å¢ãæ£è
卿ä¸çæ¨¡åæ»è§ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/SepsisAgent" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/SepsisAgent</strong><span>ICU sepsis treatment recommendation with Clinical World Modelã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/Agentic-Hospital" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/Agentic-Hospital</strong><span>Hospital-scale actionable environment for autonomous agentsã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/TwinHospital" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/TwinHospital</strong><span>HIS data-flow-driven hospital process simulationã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/OpenClaw-Medical-Skills" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/OpenClaw-Medical-Skills</strong><span>é¢å medical agents ç弿¾æè½åºã</span></span></a> </div> </details> </section> <section class="project-program-section project-theme-medical" id="dental-foundation-models" data-project-tags="Healthcare AI;Foundation Models;Evaluation & Interpretability" data-project-url="/projects/dental-foundation-models/"> <div class="project-program-heading"> <div class="project-eyebrow">Dental LLMs, multimodal diagnosis, clinical reasoning, and benchmarks</div> <h2>çç§å¤§æ¨¡å</h2> <p> çç§æ¯å¯ä»¥ç¬ç«ææ¹åçåç´å»ç AIï¼æ¢æ DentalGPT è¿æ ·ç夿¨¡æçç§å¤§æ¨¡åï¼ä¹æ GlobalDentBenchãOmniDentBench çé¢åä¸´åºæ¨çãå®å
¨é£é©åä¸å®¶æ ¡åç benchmarkãå®å¼ºè°å£è
å½±åãç
便¨çãä¸ç§ taxonomyãçå®ä¸´åºé£é©åä¸å®¶è¯åã </p> <div class="project-actions"> <a class="project-action primary" href="/projects/dental-foundation-models/">Open Dental LLMs</a> <a class="project-action" href="https://freedomdle.cn/" target="_blank" rel="noopener">Open OmniDentBench</a> <a class="project-action" href="https://github.com/FreedomIntelligence/DentalGPT" target="_blank" rel="noopener">Open DentalGPT</a> </div> </div> <div class="project-program-layout"> <div class="project-program-media"> <img src="/assets/img/media-roundup-2026/cmb-overview.png" alt="Dental foundation models and benchmarks"> </div> <div> <h3>ç¸å
³è®ºæä¸ benchmark</h3> <ol class="project-paper-list"> <li> <a href="https://arxiv.org/abs/2512.11558" target="_blank" rel="noopener">DentalGPT</a>
1ï¼ç¨ 120K+ çç§å¾åãdomain knowledge injection å GRPO 强å夿¨¡æå¤ææ¨çã</li> <li> <a href="https://arxiv.org/abs/2605.24636" target="_blank" rel="noopener">GlobalDentBench</a>ï¼è¦ç 88 个å½å®¶/å°åºã14 个çç§ä¸ç§ã8,978 éä¸å®¶æ ¡åé¢ç multinational dental benchmarkã</li> <li> <a href="https://freedomdle.cn/" target="_blank" rel="noopener">OmniDentBench</a>ï¼é¢å夿䏴åºå³çåçç©å»å¦ç ç©¶çå
¨ççç§åºåè¯æµå¹³å°ã</li> <li> <a href="https://arxiv.org/abs/2508.20416" target="_blank" rel="noopener">DentalBench</a>ï¼åè¯çç§ QA benchmark åå¤§è§æ¨¡ DentalCorpusï¼ç¨äºè¯ä¼°åå¢å¼ºçç§çè§£ã</li> <li> <a href="https://arxiv.org/abs/2511.22055" target="_blank" rel="noopener">OralGPT-Omni</a>ï¼çç§å¤æ¨¡æå¤§æ¨¡åä¸ MMOral-Uni benchmarkï¼è¦çå¤ç§å£è
å½±åä»»å¡ã</li> </ol> </div> </div> <details class="project-resources-panel"> <summary><span>模å / benchmark / å¹³å°</span><strong>Dental AI stack</strong></summary> <div class="project-stack project-resource-stack"> <a class="project-stack-row" href="/projects/dental-foundation-models/"><span class="project-stack-type">Project</span><span class="project-stack-main"><strong>Dental Foundation Models</strong><span>çç§å¤§æ¨¡åãçç§ benchmark å OmniDentBench æ»è§ã</span></span></a> <a class="project-stack-row" href="https://freedomdle.cn/" target="_blank" rel="noopener"><span class="project-stack-type">Platform</span><span class="project-stack-main"><strong>OmniDentBench</strong><span>å
¨ççç§åºåè¯æµå¹³å°å ODB leaderboardã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/DentalGPT" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/DentalGPT</strong><span>çç§å¤æ¨¡æå¤§æ¨¡åãè®ç» pipeline åè¯æµè®¾ç½®ã</span></span></a> <a class="project-stack-row" href="https://github.com/FreedomIntelligence/GlobalDentBench" target="_blank" rel="noopener"><span class="project-stack-type">GitHub</span><span class="project-stack-main"><strong>FreedomIntelligence/GlobalDentBench</strong><span>ä¸å®¶æ ¡åçè·¨å½çç§ä¸´åºæ¨ç benchmarkã</span></span></a> <a class="project-stack-row" href="https://huggingface.co/datasets/FreedomIntelligence/GlobalDentBench-OA" target="_blank" rel="noopener"><span class="project-stack-type">HF Dataset</span><span class="project-stack-main"><strong>GlobalDentBench-OA</strong><span>GlobalDentBench 弿¾è¯æµåéã</span></span></a> </div> </details> </section> </div>
1<script>document.addEventListener("DOMContentLoaded",function(){function t(t){return c[t]||t}function e(t){return(t.dataset.projectTags||"").split(";").map(function(t){return t.trim()}).filter(Boolean)}function n(n,i){n=t(n);let c=0;if(r.forEach(function(t){const o="all"===n||e(t).includes(n);t.classList.toggle("is-hidden",!o),o&&(c+=1)}),o.forEach(function(t){const e=t.dataset.projectFilter===n;t.classList.toggle("is-active",e),t.setAttribute("aria-pressed",String(e))}),a&&(a.hidden=c>0),i){const t=new URL(window.location.href);"all"===n?t.searchParams.delete("tag"):t.searchParams.set("tag",n),window.history.replaceState({},"",t)}}const o=Array.from(document.querySelectorAll("[data-project-filter]")),r=Array.from(document.querySelectorAll(".project-program-section")),a=document.querySelector(".project-empty"),i="LLM Agents, Applications & Human-Agent Interaction",c={"LLM Agents and Applications":i,"Human-Agent Interaction":i,"Clinical Agents & Hospital World Models":"Clinical & Hospital World Models"};r.forEach(function(t){const e=t.dataset.projectUrl;if(!e)return;const n=t.querySelector("h2");t.classList.add("is-clickable"),t.setAttribute("role","link"),t.setAttribute("tabindex","0"),n&&t.setAttribute("aria-label","Open project: "+n.textContent.trim()),t.addEventListener("click",function(t){const n=t.target;n instanceof Element&&(n.closest('a, button, summary, details, input, select, textarea, [role="button"], .project-resources-panel')||(window.location.href=e))}),t.addEventListener("keydown",function(t){"Enter"===t.key&&(window.location.href=e)})}),o.forEach(function(t){t.addEventListener("click",function(){n(t.dataset.projectFilter,!0)})});const s=t(new URLSearchParams(window.location.search).get("tag"));n(o.some(function(t){return t.dataset.projectFilter===s})?s:"all",!1)});</script>
1 </article> </div> </div> <footer class="sticky-bottom mt-5"> <div class="container"> © Copyright 2026 Freedom AI . 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>. </div> </footer>
1<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/jquery.min.js" integrity="sha256-/xUj+3OJU5yExlq6GSYGSHk7tPXikynS7ogEvDej/m4=" crossorigin="anonymous"></script>
1
1<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/js/bootstrap.bundle.min.js" integrity="sha256-fgLAgv7fyCGopR/gBNq2iW3ZKIdqIcyshnUULC4vex8=" crossorigin="anonymous"></script>
1
1<script src="https://cdn.jsdelivr.net/npm/[email protected]/js/mdb.min.js" integrity="sha256-NdbiivsvWt7VYCt6hYNT3h/th9vSTL4EDWeGs5SN3DA=" crossorigin="anonymous"></script>
1
1<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/masonry.pkgd.min.js" integrity="sha256-Nn1q/fx0H7SNLZMQ5Hw5JLaTRZp0yILA/FRexe19VdI=" crossorigin="anonymous"></script>
1
1<script defer src="https://cdn.jsdelivr.net/npm/imagesloaded@4/imagesloaded.pkgd.min.js"></script>
1
1<script defer src="/assets/js/masonry.js" type="text/javascript"></script>
1
1<script defer src="/assets/js/common.js"></script>
1
1<script defer src="/assets/js/language-toggle.js"></script>
1
1<script type="text/javascript">window.MathJax={tex:{inlineMath:[["$","$"]],tags:"ams"}};</script>
1
1<script defer type="text/javascript" id="MathJax-script" src="https://cdn.jsdelivr.net/npm/[email protected]/es5/tex-mml-chtml.js"></script>
1 </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.