Hugging Face Daily Papers · June 26, 2026 · 4 min read

OPID: On-Policy Skill Distillation for Agentic Reinforcement Learning

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

<a href=\"https://cdn-uploads.huggingface.co/production/uploads/6747de57f8cab58c22ec94a2/qDEGap3JRvtGgPffHWhQh.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/6747de57f8cab58c22ec94a2/qDEGap3JRvtGgPffHWhQh.png\" alt=\"OPID_framework_01\"></a></p>\n","updatedAt":"2026-06-26T02:16:45.547Z","author":{"_id":"6747de57f8cab58c22ec94a2","avatarUrl":"/avatars/5bae0341862fac24564781c0fa32aac5.svg","fullname":"Jinyang Wu","name":"Jinyang23","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":11,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.27833470702171326},"editors":["Jinyang23"],"editorAvatarUrls":["/avatars/5bae0341862fac24564781c0fa32aac5.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.26790","authors":[{"_id":"6a3de03c3b43e283349ec169","name":"Shuo Yang","hidden":false},{"_id":"6a3de03c3b43e283349ec16a","user":{"_id":"6747de57f8cab58c22ec94a2","avatarUrl":"/avatars/5bae0341862fac24564781c0fa32aac5.svg","isPro":false,"fullname":"Jinyang Wu","user":"Jinyang23","type":"user","name":"Jinyang23"},"name":"Jinyang Wu","status":"claimed_verified","statusLastChangedAt":"2026-06-27T15:23:43.228Z","hidden":false},{"_id":"6a3de03c3b43e283349ec16b","user":{"_id":"676127cf11b19ea602bb202a","avatarUrl":"/avatars/dfd802a24bd63e509728159ebb1769f6.svg","isPro":false,"fullname":"Zhengxi Lu","user":"LZXzju","type":"user","name":"LZXzju"},"name":"Zhengxi Lu","status":"claimed_verified","statusLastChangedAt":"2026-06-27T15:23:40.687Z","hidden":false},{"_id":"6a3de03c3b43e283349ec16c","name":"Yuhao Shen","hidden":false},{"_id":"6a3de03c3b43e283349ec16d","name":"Fan Zhang","hidden":false},{"_id":"6a3de03c3b43e283349ec16e","name":"Lang Feng","hidden":false},{"_id":"6a3de03c3b43e283349ec16f","name":"Shuai Zhang","hidden":false},{"_id":"6a3de03c3b43e283349ec170","name":"Haoran Luo","hidden":false},{"_id":"6a3de03c3b43e283349ec171","name":"Zheng Lian","hidden":false},{"_id":"6a3de03c3b43e283349ec172","name":"Zhengqi Wen","hidden":false},{"_id":"6a3de03c3b43e283349ec173","name":"Jianhua Tao","hidden":false}],"publishedAt":"2026-06-25T00:00:00.000Z","submittedOnDailyAt":"2026-06-26T00:00:00.000Z","title":"OPID: On-Policy Skill Distillation for Agentic Reinforcement Learning","submittedOnDailyBy":{"_id":"6747de57f8cab58c22ec94a2","avatarUrl":"/avatars/5bae0341862fac24564781c0fa32aac5.svg","isPro":false,"fullname":"Jinyang Wu","user":"Jinyang23","type":"user","name":"Jinyang23"},"summary":"Outcome-based reinforcement learning provides a stable optimization backbone for language agents, but its sparse trajectory-level rewards provide little guidance on which intermediate decisions should be reinforced or suppressed. On-policy self-distillation offers dense token-level supervision, yet existing skill-conditioned variants often rely on external skill memories or retrieved privileged context, which are costly to maintain and can be mismatched with the state distribution induced by the current policy in multi-turn interaction. We propose OPID (On-Policy Skill Distillation), a framework that extracts skill supervision directly from completed on-policy trajectories. OPID represents trajectory hindsight as hierarchical skills: episode-level skills capture global workflows or failure-avoidance rules, while step-level skills capture local decision knowledge at critical timesteps. A critical-first routing mechanism uses step-level skills when critical decisions are identified and falls back to episode-level skills as default guidance otherwise. The selected skill is injected into the interaction history, allowing the old policy to re-score the same sampled response under both original and skill-augmented contexts. The resulting log-probability shift yields a token-level self-distillation advantage, which is combined with the outcome advantage for policy optimization. OPID thus preserves RL as the primary training objective while introducing dense, distribution-matched hindsight supervision. Experiments on ALFWorld, WebShop and Search-based QA demonstrate that OPID generally improves agent performance, sample efficiency, and robustness over outcome-only RL and existing skill-distillation baselines. Our code is available at https://github.com/jinyangwu/OPID/tree/main.","upvotes":40,"discussionId":"6a3de03d3b43e283349ec174","githubRepo":"https://github.com/jinyangwu/OPID","githubRepoAddedBy":"user","ai_summary":"On-policy skill distillation framework extracts dense hindsight supervision from completed trajectories to improve language agent training efficiency and performance.","ai_keywords":["reinforcement learning","self-distillation","skill-conditioned variants","on-policy trajectories","hierarchical skills","critical-first routing","token-level supervision","policy optimization","outcome-based reinforcement learning"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":35},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"676127cf11b19ea602bb202a","avatarUrl":"/avatars/dfd802a24bd63e509728159ebb1769f6.svg","isPro":false,"fullname":"Zhengxi Lu","user":"LZXzju","type":"user"},{"_id":"695f5bad4938eb7e6b3f9693","avatarUrl":"/avatars/de7693b683158e80186eba91a7d1f1ab.svg","isPro":false,"fullname":"Xintian Li","user":"LiixT","type":"user"},{"_id":"652f5c12c22d404ebfa126d2","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/652f5c12c22d404ebfa126d2/SoG5VlBelL3YkZZ41b1_o.jpeg","isPro":false,"fullname":"Zheng Lian","user":"zeroqiaoba","type":"user"},{"_id":"6639ad487c0ab4fd9df1dde5","avatarUrl":"/avatars/8cc99f6ed8f8c1b2a14dde797a991a8c.svg","isPro":false,"fullname":"Fan Zhang","user":"Karl28","type":"user"},{"_id":"6747de57f8cab58c22ec94a2","avatarUrl":"/avatars/5bae0341862fac24564781c0fa32aac5.svg","isPro":false,"fullname":"Jinyang Wu","user":"Jinyang23","type":"user"},{"_id":"6936bb025726164a40ff8ee1","avatarUrl":"/avatars/86079cbf21d7120994bc28b65b44840f.svg","isPro":false,"fullname":"JIELONG YAN","user":"jasonvyan","type":"user"},{"_id":"67cff17e63ae74271047b474","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/u_moerdIJmtitopcvAzEc.png","isPro":false,"fullname":"Kevin Xue","user":"JingsiAI","type":"user"},{"_id":"65b310e48b7c508c8e6f92c0","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65b310e48b7c508c8e6f92c0/RLAmKMfMFbwE9mmi8O381.png","isPro":false,"fullname":"ZhouQing","user":"mrazhou","type":"user"},{"_id":"66ba29dd59e8e7a957154c5f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66ba29dd59e8e7a957154c5f/VvVS7IZNPUIB023GAEf5u.png","isPro":false,"fullname":"Lang Feng","user":"langfeng01","type":"user"},{"_id":"6736dd067506985f27c236c7","avatarUrl":"/avatars/7900b72ee3b446ee73fc4ffe7cae97ac.svg","isPro":false,"fullname":"lucas Zhai","user":"lucaszgc","type":"user"},{"_id":"68e2740471512dfe660dff46","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/6BcKV0oL0LD9jgVSi00Oz.png","isPro":false,"fullname":"xnchang","user":"LGXNBXN","type":"user"},{"_id":"674d9fe27519b842ae38d604","avatarUrl":"/avatars/9d5b0b473e871154c30669cc0d1e121b.svg","isPro":false,"fullname":"lyj","user":"hardworkinglyj","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.26790.md","query":{}}">

Papers

arxiv:2606.26790

OPID: On-Policy Skill Distillation for Agentic Reinforcement Learning

Published on Jun 25

· Submitted by

Jinyang Wu on Jun 26

Upvote

Authors:

Jinyang Wu ,

Zhengxi Lu ,

Abstract

On-policy skill distillation framework extracts dense hindsight supervision from completed trajectories to improve language agent training efficiency and performance.

Generated by Qwen/Qwen2.5-Coder-32B-Instruct

Outcome-based reinforcement learning provides a stable optimization backbone for language agents, but its sparse trajectory-level rewards provide little guidance on which intermediate decisions should be reinforced or suppressed. On-policy self-distillation offers dense token-level supervision, yet existing skill-conditioned variants often rely on external skill memories or retrieved privileged context, which are costly to maintain and can be mismatched with the state distribution induced by the current policy in multi-turn interaction. We propose OPID (On-Policy Skill Distillation), a framework that extracts skill supervision directly from completed on-policy trajectories. OPID represents trajectory hindsight as hierarchical skills: episode-level skills capture global workflows or failure-avoidance rules, while step-level skills capture local decision knowledge at critical timesteps. A critical-first routing mechanism uses step-level skills when critical decisions are identified and falls back to episode-level skills as default guidance otherwise. The selected skill is injected into the interaction history, allowing the old policy to re-score the same sampled response under both original and skill-augmented contexts. The resulting log-probability shift yields a token-level self-distillation advantage, which is combined with the outcome advantage for policy optimization. OPID thus preserves RL as the primary training objective while introducing dense, distribution-matched hindsight supervision. Experiments on ALFWorld, WebShop and Search-based QA demonstrate that OPID generally improves agent performance, sample efficiency, and robustness over outcome-only RL and existing skill-distillation baselines. Our code is available at https://github.com/jinyangwu/OPID/tree/main.