50%.\n\n💻 Code: https://github.com/ZJU-REAL/Relay-OPD\n🌐 Project: https://zju-real.github.io/Relay-OPD/","html":"<p>We introduce Relay-OPD 🏃 — on-policy distillation that fixes prefix failure: once a student commits to a wrong reasoning direction early, the entire rollout builds on the mistake, yielding unreliable supervision and wasted compute.</p>\n<p>Key observation: on failed prefixes, teacher and student diverge — the teacher wants to stop and redirect (\"But\", \"Wait\"…), while the student presses on. This divergence is observable online, with no verifier, reward model, or labels, and it marks exactly where to intervene. Relay-OPD lets the teacher briefly take over at these handoff triggers, then pass the baton back to the student, with a relay budget keeping intervention early and local. The entire rollout runs in a single speculative-decoding engine (student drafts, teacher verifies).</p>\n<p>Results: with a Qwen3-4B teacher and 0.6B/1.7B students on 8 math reasoning benchmarks: +5.73% avg. over standard OPD, +1.49% over the strongest baseline FastOPD, while cutting training trajectory length by &gt;50%.</p>\n<p>💻 Code: <a href=\"https://github.com/ZJU-REAL/Relay-OPD\" rel=\"nofollow\">https://github.com/ZJU-REAL/Relay-OPD</a><br>🌐 Project: <a href=\"https://zju-real.github.io/Relay-OPD/\" rel=\"nofollow\">https://zju-real.github.io/Relay-OPD/</a></p>\n","updatedAt":"2026-07-29T08:07:29.235Z","author":{"_id":"6692aff88db712bad780f02a","avatarUrl":"/avatars/5dc4b1c27c70f6a64864711dbff4910f.svg","fullname":"xhl","name":"zjuxhl","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":5,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.876446545124054},"editors":["zjuxhl"],"editorAvatarUrls":["/avatars/5dc4b1c27c70f6a64864711dbff4910f.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.26057","authors":[{"_id":"6a699d6c9d3a1231d492b92e","name":"Haolei Xu","hidden":false},{"_id":"6a699d6c9d3a1231d492b92f","name":"Xiaowen Xu","hidden":false},{"_id":"6a699d6c9d3a1231d492b930","name":"Haiwen Hong","hidden":false},{"_id":"6a699d6c9d3a1231d492b931","name":"Zixuan Ni","hidden":false},{"_id":"6a699d6c9d3a1231d492b932","name":"Hongxing Li","hidden":false},{"_id":"6a699d6c9d3a1231d492b933","name":"Yiwen Qiu","hidden":false},{"_id":"6a699d6c9d3a1231d492b934","name":"Weiming Lu","hidden":false},{"_id":"6a699d6c9d3a1231d492b935","name":"Yongliang Shen","hidden":false}],"publishedAt":"2026-07-28T00:00:00.000Z","submittedOnDailyAt":"2026-07-29T00:00:00.000Z","title":"Pass the Baton: Trajectory-Relayed On-Policy Distillation","submittedOnDailyBy":{"_id":"6692aff88db712bad780f02a","avatarUrl":"/avatars/5dc4b1c27c70f6a64864711dbff4910f.svg","isPro":false,"fullname":"xhl","user":"zjuxhl","type":"user","name":"zjuxhl"},"summary":"On-policy distillation (OPD) grounds token-level supervision in the student's own trajectory, yet suffers from prefix failure: once the student commits to a wrong reasoning direction, all subsequent generation builds on this deviation, producing misdirected continuations that elicit unreliable supervision and waste compute. We identify a teacher-student continuation asymmetry on failed prefixes, where the teacher tends to redirect while the student continues along the original direction, and convert it into a label-free handoff trigger in Relay On-Policy Distillation (Relay-OPD). During training, Relay-OPD constructs relay trajectories by letting the teacher briefly take over at detected trigger points to produce a teacher leg, after which the student resumes and is optimized on the resulting trajectory. A limited relay budget concentrates intervention on critical early positions while limiting departure from the student policy. With a Qwen3-4B-Instruct-2507 teacher and Qwen3-0.6B/1.7B-Non-Thinking students on eight mathematical reasoning benchmarks, Relay-OPD achieves the best or second-best results on every benchmark, outperforming standard OPD by +5.73% and the strongest baseline FastOPD by +1.49% on average for 1.7B, with consistent gains at 0.6B. Training trajectory length is reduced by over 50%.","upvotes":13,"discussionId":"6a699d6c9d3a1231d492b936","projectPage":"https://zju-real.github.io/Relay-OPD/","githubRepo":"https://github.com/ZJU-REAL/Relay-OPD","githubRepoAddedBy":"user","githubStars":3,"organization":{"_id":"61bac2af530e5c78d7b99667","name":"zju","fullname":"Zhejiang University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/5e1058e9fcf41d740b69966d/7G1xjlxwCdMEmKcxNR0n5.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"67505c291372c3446d0a1f5d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/rOVvWED4Fm6e8aWJDP5pj.jpeg","isPro":false,"fullname":"Hongxing Li","user":"hongxingli","type":"user"},{"_id":"67543820c3af453d7b3e1d5e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/67543820c3af453d7b3e1d5e/RbAZ9AQlxpy5E5Is-QN8b.jpeg","isPro":false,"fullname":"Dingming Li","user":"lidingm","type":"user"},{"_id":"676127cf11b19ea602bb202a","avatarUrl":"/avatars/dfd802a24bd63e509728159ebb1769f6.svg","isPro":false,"fullname":"Zhengxi Lu","user":"LZXzju","type":"user"},{"_id":"677d3523a6918748bf81d0e9","avatarUrl":"/avatars/c9961ac54d089efb36db20c421c2bea2.svg","isPro":false,"fullname":"wangzixuan","user":"wangzx1210","type":"user"},{"_id":"6692aff88db712bad780f02a","avatarUrl":"/avatars/5dc4b1c27c70f6a64864711dbff4910f.svg","isPro":false,"fullname":"xhl","user":"zjuxhl","type":"user"},{"_id":"66d2c8a53c8270397e4dba96","avatarUrl":"/avatars/bc7240ce8b03c19cee3b0f47c87fd616.svg","isPro":false,"fullname":"Ruijie Jian","user":"Jerryton2003","type":"user"},{"_id":"684e69b194781ed88871b1a6","avatarUrl":"/avatars/d150e7668f8e1be9fff2bcdcbff7dc21.svg","isPro":false,"fullname":"zyx","user":"zyx221","type":"user"},{"_id":"6875b4d439e32e284562f53f","avatarUrl":"/avatars/e234beeee37c649ef07f5a3e5f634441.svg","isPro":false,"fullname":"W.Jimmy","user":"WJimmy","type":"user"},{"_id":"6604160b8d50be909bea5ff6","avatarUrl":"/avatars/04c17cbe98fec21aec73bb6e208cb4e8.svg","isPro":false,"fullname":"Jinhao Chen","user":"DexterChan","type":"user"},{"_id":"63a942dd2e05ca32e35335df","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63a942dd2e05ca32e35335df/kuKfBLEXfWnvnoUUmoXW6.jpeg","isPro":false,"fullname":"haoming xu","user":"haomingx","type":"user"},{"_id":"6485bd278d14bcd5cdbb7c8d","avatarUrl":"/avatars/1427cf1a72b5db0cb263ad45885cf925.svg","isPro":false,"fullname":"Wenqi Zhang","user":"zwq2018","type":"user"},{"_id":"65ef2d78e26bcf263dc7a806","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65ef2d78e26bcf263dc7a806/3QSx6Yk_thl7YARek5sx4.png","isPro":false,"fullname":"Fan Yuan","user":"Leoyfan","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"61bac2af530e5c78d7b99667","name":"zju","fullname":"Zhejiang University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/5e1058e9fcf41d740b69966d/7G1xjlxwCdMEmKcxNR0n5.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.26057.md","query":{}}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2607.26057 in a model README.md to link it from this page.

Cite arxiv.org/abs/2607.26057 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2607.26057 in a Space README.md to link it from this page.

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