Code: <a href=\"https://github.com/YuYingLi0/FiRe-OPD\" rel=\"nofollow\">https://github.com/YuYingLi0/FiRe-OPD</a></p>\n","updatedAt":"2026-06-04T04:41:48.249Z","author":{"_id":"649d54b314afbb10ce2a9eeb","avatarUrl":"/avatars/15c325d8c2273ff63569f23015e98486.svg","fullname":"Hangjie Yuan","name":"JacobYuan","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":11,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8621976971626282},"editors":["JacobYuan"],"editorAvatarUrls":["/avatars/15c325d8c2273ff63569f23015e98486.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.02684","authors":[{"_id":"6a202b6615100c5272a841d5","name":"Yuying Li","hidden":false},{"_id":"6a202b6615100c5272a841d6","name":"Leqi Zheng","hidden":false},{"_id":"6a202b6615100c5272a841d7","name":"Yongzi Yu","hidden":false},{"_id":"6a202b6615100c5272a841d8","name":"Wenrui Zhou","hidden":false},{"_id":"6a202b6615100c5272a841d9","name":"Xuchang Zhong","hidden":false},{"_id":"6a202b6615100c5272a841da","name":"Xing Hu","hidden":false},{"_id":"6a202b6615100c5272a841db","name":"Jing Jin","hidden":false},{"_id":"6a202b6615100c5272a841dc","name":"Huangjie Yuan","hidden":false},{"_id":"6a202b6615100c5272a841dd","name":"Tao Feng","hidden":false}],"publishedAt":"2026-06-01T00:00:00.000Z","submittedOnDailyAt":"2026-06-04T00:00:00.000Z","title":"Filter, Then Reweight: Rethinking Optimization Granularity in On-Policy Distillation","submittedOnDailyBy":{"_id":"649d54b314afbb10ce2a9eeb","avatarUrl":"/avatars/15c325d8c2273ff63569f23015e98486.svg","isPro":false,"fullname":"Hangjie Yuan","user":"JacobYuan","type":"user","name":"JacobYuan"},"summary":"On-Policy distillation (OPD) in large language models is shifting from full-trace KL supervision toward more selective training paradigms. Recent OPD methods increasingly focus on selecting which trajectories to learn from, which tokens are most informative, and which supervision signals are most reliable. Motivated by this trend, we rethink optimization granularity of OPD and propose \\fireicon\\ FiRe-OPD (Filter, then Reweight), which jointly adjusts supervision signals at both trajectory and token levels. In details, FiRe-OPD first filters trajectories to remove low-quality rollout samples, and then applies soft reweighting within the retained trajectories to emphasize informative tokens. Compared with hard token selection, FiRe-OPD leverages a soft-weighting mechanism to effectively mitigate information loss and enhance optimization stability, thereby achieving finer-grained OPD optimization. We validate the effectiveness of FiRe-OPD across strong-to-weak, single-teacher, and multi-teacher settings, and demonstrate its superiority over recent token-level OPD methods ( (e.g., +6.25 on AIME 2024 in strong-to-weak, +18.81 on Miner in multi-teacher). Our code is available at https://github.com/YuYingLi0/FiRe-OPD.","upvotes":8,"discussionId":"6a202b6615100c5272a841f0","githubRepo":"https://github.com/YuYingLi0/FiRe-OPD","githubRepoAddedBy":"user","ai_summary":"FiRe-OPD improves on-policy distillation in large language models by filtering low-quality trajectories and applying soft reweighting to enhance informative token selection and optimization stability.","ai_keywords":["on-policy distillation","KL supervision","trajectory selection","token selection","supervision signals","soft reweighting","optimization granularity","FiRe-OPD","trajectory filtering","token weighting"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":4,"organization":{"_id":"628735cbc83a2d6ab8d14a66","name":"Tsinghua","fullname":"Tsinghua University","avatar":"https://www.gravatar.com/avatar/6c5c1441e3283e7543342e59277ea219?d=retro&size=100"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"649d54b314afbb10ce2a9eeb","avatarUrl":"/avatars/15c325d8c2273ff63569f23015e98486.svg","isPro":false,"fullname":"Hangjie Yuan","user":"JacobYuan","type":"user"},{"_id":"64f44021220c2e5e96628595","avatarUrl":"/avatars/43bcb389760d1037dde1f81c4398d260.svg","isPro":true,"fullname":"gdwind LQ","user":"gdwind","type":"user"},{"_id":"644d2532d185572dd1e48f90","avatarUrl":"/avatars/5831acebb02d8bc8f80f56b7b11c7c69.svg","isPro":false,"fullname":"Zhu","user":"zzzhu","type":"user"},{"_id":"680b3903f209ad3256527538","avatarUrl":"/avatars/23d715718f23d00cc8c370e2e21c514e.svg","isPro":false,"fullname":"XuzhaoLi","user":"LXZ1OOO","type":"user"},{"_id":"678609789a285d232ee14157","avatarUrl":"/avatars/a6cb2c571d9ef6deb0b1659f754afe7f.svg","isPro":false,"fullname":"Weichu Xie","user":"akarinmoe","type":"user"},{"_id":"6708cca079b2d9e035c4a54e","avatarUrl":"/avatars/ed7daf61627ea910ec2de4582abf5a0d.svg","isPro":false,"fullname":"xichen","user":"xichen-fy","type":"user"},{"_id":"66a246cae22bfd8d72560696","avatarUrl":"/avatars/97cec177a0eb0a8465e3532b888ef1d4.svg","isPro":false,"fullname":"zrchen","user":"zrchen03","type":"user"},{"_id":"69bcc905518f6d1f3d2b86df","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/t-ZEqYEkqPI8-saats0nT.jpeg","isPro":false,"fullname":"Liu Jingyi","user":"gao-wenxuan2","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"628735cbc83a2d6ab8d14a66","name":"Tsinghua","fullname":"Tsinghua University","avatar":"https://www.gravatar.com/avatar/6c5c1441e3283e7543342e59277ea219?d=retro&size=100"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.02684.md"}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2606.02684 in a model README.md to link it from this page.

Cite arxiv.org/abs/2606.02684 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2606.02684 in a Space README.md to link it from this page.

Add this paper to a collection to link it from this page.