[{"data":1,"prerenderedAt":77},["ShallowReactive",2],{"profile-572942446ad9cf9201b83cfe123c7162":3,"works-572942446ad9cf9201b83cfe123c7162":22,"scholar-stats":54},{"id":4,"name":5,"avatar":6,"org":7,"title":8,"bio":9,"intro":10,"badgeNo":11,"scholarUrl":12,"emailPublic":13,"github":14,"zhihu":11,"xhs":11,"experience":15,"invitedAt":11,"status":16,"directions":17,"works":21},"572942446ad9cf9201b83cfe123c7162","崔涵","https:\u002F\u002F6461-dayan-ai4s-d2gr3v6bc05b0eccd-1304768804.tcb.qcloud.la\u002Fweda-uploader\u002Ff8e0805d4fbdb62256dfd7cb66cf605f-%E5%B4%94%E6%B6%B5.jpg","西湖大学","博士研究生","关注于模型后训练和蒸馏技术","我是西湖大学的博士生。此前，我在哈尔滨工业大学获得了计算机科学学士和硕士学位。我的研究重点是大型语言模型，特别是推理、强化学习和后训练，主要关注在了解语言模型是如何获得和提高其推理能力，以及开发更有效和可解释的方法来训练它们。","","https:\u002F\u002Fhancui.site","cuihan@westlake.edu.cn","https:\u002F\u002Fgithub.com\u002FHancCui","2025.08-现在 西湖大学 博士生在读\n2022.08–2024.12 哈尔滨工业大学 硕士研究生\n2018.08–2022.06 哈尔滨工业大学 本科生","on",[18,19,20],"大语言模型","强化学习","模型后训练",[],[23,34,44],{"id":24,"profileId":4,"type":25,"title":26,"authors":27,"venue":28,"date":29,"tags":30,"intro":32,"image":11,"url":11,"sort":33,"status":16},"a2453f646ad9d42405059063745f9d5c","paper","Direct value optimization: Improving chain-of-thought reasoning in llms with refined values","Hongbo Zhang, Han Cui, Guangsheng Bao, Linyi Yang, Jun Wang, Y Zhang","EMNLP2025","2025.11",[19,31],"价值模型","\u003Cp>\u003Cspan style=\"font-family:PingFangSC-Regular, 苹方, sans-serif\">\u003Cspan style=\"font-size:14px\">\u003Cspan style=\"color:#344054\">基于MCTS估计中间步骤的价值，并将其做为模型训练的监督信号。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>",1,{"id":35,"profileId":4,"type":25,"title":36,"authors":37,"venue":38,"date":39,"tags":40,"intro":42,"image":11,"url":11,"sort":43,"status":16},"ff49d0d36ad9d479018ce6bb715f030e","Gains and Collapse in On-Policy Distillation: A Reinforcement Learning Perspective","Han Cui, Jianhao Yan, Yun Luo, Hongbo Zhang, Zhizhang Fu, Yyue Zhang","ICLR27在投","2026.09",[19,41],"在线蒸馏","\u003Cp>\u003Cspan style=\"font-family:PingFangSC-Regular, 苹方, sans-serif\">\u003Cspan style=\"font-size:14px\">\u003Cspan style=\"color:#344054\">从强化学习的视角分析了OPD的成功训练中性能提升的来源和崩溃训练中的原因。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>",2,{"id":45,"profileId":4,"type":25,"title":46,"authors":47,"venue":48,"date":49,"tags":50,"intro":52,"image":11,"url":11,"sort":53,"status":16},"ff49d0d36ad9d500018cf2c103f80c17","Exploring hybrid question answering via program-based prompting","Qi Shi, Han Cui, Haofeng Wang, Qingfu Zhu, Wanxiang Che, Ting Liu","ACL2024","2024.08",[51],"异构数据推理","\u003Cp>\u003Cspan style=\"font-family:PingFangSC-Regular, 苹方, sans-serif\">\u003Cspan style=\"font-size:14px\">\u003Cspan style=\"color:#344054\">通过异构数据转换为结构化的数据结构，让LLM基于代码在异构数据上实现更高效，更准确的推理过程。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>",3,{"views":55,"hotIds":73},{"853d22036a97e9a3017cb879318fbfa9":56,"a8add1456a96c24d012f36ee3a3e6e80":57,"a8add1456a96c24d012f36ef55b51133":58,"a8add1456a96c24d012f36f02ef5e9b4":59,"a8add1456a96c24d012f36f14e4da6e6":59,"a8add1456a96c24d012f36f2235a9643":60,"a8add1456a96c24d012f36f364c51aca":59,"a8add1456a96c24d012f36f43c5b2575":61,"a8add1456a96c24d012f36f5055a7687":62,"a2453f646ad864b9048cd5ea3c0a4cdf":63,"a2453f646ad864b9048cd5eb700f1c13":64,"572942446ad98b2501b418a57fd29780":65,"572942446ad98b2501b418a968974e47":66,"572942446ad98b2501b418a708985667":53,"572942446ad98b2501b418a866ceb5e8":67,"572942446ad9cf9201b83d000821e9dc":68,"572942446ad9cf9201b83cfe123c7162":69,"572942446ad9cf9201b83cff5362a431":70,"fa564f266ae1e2390042b77e034d91fd":71,"255db4156af2c484002d80454c8c2b7d":72,"255db4156af417090039154a0ff49242":33},106,92,31,13,11,12,22,10,57,98,88,4,46,36,51,38,16,[74,75,76],"572942446ad98b2501b418a968974e47","572942446ad98b2501b418a57fd29780","fa564f266ae1e2390042b77e034d91fd",1794993163111]