We are excited to share <strong>Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training</strong>.</p>\n<p>Past post-training successes do not necessarily remain valid after the parent model, data, or training stage changes. We formulate this challenge as <strong>conditional experience transfer</strong> and introduce <strong>Boundary-Calibrated Intervention Transfer (BCIT)</strong>, a transparent <code>Reject / Validate / Train</code> policy that binds update evidence to its source context, vetoes hard conflicts, and runs bounded current-parent trials when applicability is uncertain.</p>\n<p>Across finance reasoning, text-to-SQL, and function calling adaptation with Qwen3-4B, BCIT achieves the highest equal-budget final-model quality among the evaluated shared-model alternatives. It improves the mean task score by <strong>2.63 points</strong> over Flat-Additive (95% CI [2.10, 3.16]) while satisfying the prespecified retention margins.</p>\n<p>📄 Paper: <a href=\"https://arxiv.org/abs/2608.26730\" rel=\"nofollow\">https://arxiv.org/abs/2608.26730</a></p>\n","updatedAt":"2026-09-04T04:54:46.693Z","author":{"_id":"65ba0e4211162bf34fb93762","avatarUrl":"/avatars/05201a3041deb298a9a9e2e3346f7911.svg","fullname":"lty","name":"spongebob0715","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8160806894302368},"editors":["spongebob0715"],"editorAvatarUrls":["/avatars/05201a3041deb298a9a9e2e3346f7911.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.26730","authors":[{"_id":"6a952114073195fee51572ce","user":{"_id":"65ba0e4211162bf34fb93762","avatarUrl":"/avatars/05201a3041deb298a9a9e2e3346f7911.svg","isPro":false,"fullname":"lty","user":"spongebob0715","type":"user","name":"spongebob0715"},"name":"Tingyun Li","status":"claimed_verified","statusLastChangedAt":"2026-08-31T08:37:59.711Z","hidden":false},{"_id":"6a952114073195fee51572cf","user":{"_id":"68b691341c52b8b741ba7fd4","avatarUrl":"/avatars/efae3c632577565eb3645e64d4f725ac.svg","isPro":false,"fullname":"wenfeng feng","user":"wenfengfwf","type":"user","name":"wenfengfwf"},"name":"Wenfeng Feng","status":"claimed_verified","statusLastChangedAt":"2026-09-04T08:45:04.166Z","hidden":false},{"_id":"6a952114073195fee51572d0","name":"Weiqing Li","hidden":false},{"_id":"6a952114073195fee51572d1","name":"Abudukelimu Wuerkaixi","hidden":false},{"_id":"6a952114073195fee51572d2","name":"Guohua Liu","hidden":false},{"_id":"6a952114073195fee51572d3","name":"Yuewei Zhang","hidden":false}],"publishedAt":"2026-08-27T00:00:00.000Z","submittedOnDailyAt":"2026-09-04T00:00:00.000Z","title":"Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training","submittedOnDailyBy":{"_id":"65ba0e4211162bf34fb93762","avatarUrl":"/avatars/05201a3041deb298a9a9e2e3346f7911.svg","isPro":false,"fullname":"lty","user":"spongebob0715","type":"user","name":"spongebob0715"},"summary":"Large language models offer broad capabilities, but adapting them to evolving domains, tools, and requirements often entails repeated post-training. Autonomous systems automate parts of this process by proposing updates, training candidates, and using evaluation feedback to select subsequent proposals. As evidence accumulates, a central problem emerges: which past update evidence remains actionable after subsequent training has changed the parent model? An update's effect depends on its parent, data, and training stage. Treating past success as context-free permission can waste compute. If the resulting child is promoted, it can also degrade the subsequent training trajectory. We formulate this problem as conditional experience transfer and introduce Boundary-Calibrated Intervention Transfer (BCIT), a method that authorizes experience reuse before weight-changing training. BCIT binds an observed effect to its source context, checks applicability conditions, vetoes candidates with named hard conflicts, and obtains current-state evidence through a bounded training trial when needed. Fully trained candidates still face a shared adoption rule, and only observed events extend memory. On one 4B model adapted across finance reasoning, text-to-SQL, and function calling, candidate updates exhibit heterogeneous target and retention effects across the evaluated contexts. Under matched candidates, evidence, and compute, BCIT authorizes fewer harmful updates and attains higher equal-budget final-model quality than the evaluated alternatives. These results support treating experience authorization as a distinct problem in autonomous post-training.","upvotes":24,"discussionId":"6a952114073195fee51572d4","ai_summary":"Boundary-Calibrated Intervention Transfer selectively reuses past training evidence by checking contextual applicability and running bounded trials, reducing harmful updates and improving final model quality in autonomous post-training.","ai_keywords":["conditional experience transfer","Boundary-Calibrated Intervention Transfer","autonomous post-training","experience authorization","weight-changing training","bounded training trial"],"ai_summary_model":"thinkingmachines/Inkling-Small"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"65ba0e4211162bf34fb93762","avatarUrl":"/avatars/05201a3041deb298a9a9e2e3346f7911.svg","isPro":false,"fullname":"lty","user":"spongebob0715","type":"user"},{"_id":"68b691341c52b8b741ba7fd4","avatarUrl":"/avatars/efae3c632577565eb3645e64d4f725ac.svg","isPro":false,"fullname":"wenfeng feng","user":"wenfengfwf","type":"user"},{"_id":"644a1dbb9c340e5e1e713153","avatarUrl":"/avatars/21cb93ad067a798a39829ef7e67c70b8.svg","isPro":false,"fullname":"JGC","user":"Nothing2Say","type":"user"},{"_id":"684d57f26e04c265777ead3f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/cuOj-bQqukSZreXgUJlfm.png","isPro":false,"fullname":"Joakim Lee","user":"Reinforcement4All","type":"user"},{"_id":"63ac5701c21e60a3e9b58aa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63ac5701c21e60a3e9b58aa7/g6EX7diOpuA94R2ab-rZC.png","isPro":true,"fullname":"Dipankar Sarkar","user":"dipankarsarkar","type":"user"},{"_id":"66f6aeff51f144336529ff3c","avatarUrl":"/avatars/ddd76034f7f8c075c8223fb3a9e9c384.svg","isPro":false,"fullname":"urkax","user":"urkax","type":"user"},{"_id":"6284d00fb6ba01dcb0c3869f","avatarUrl":"/avatars/eb21d049fd47e1f5ffcc080709b18f32.svg","isPro":false,"fullname":"liweiqing","user":"lwq","type":"user"},{"_id":"66bd9d3385c8bcb5f2a29f8c","avatarUrl":"/avatars/c630303c8eda8d29abe22c66c4b787f3.svg","isPro":false,"fullname":"SongJingyi","user":"jennsoo","type":"user"},{"_id":"6645cef915c233efdf1ef9e6","avatarUrl":"/avatars/2d5dc7933f10c61e5e15810efbd9c732.svg","isPro":false,"fullname":"yuewei zhang","user":"hugging-llm","type":"user"},{"_id":"68d6a2fcc6e9f4afa786573c","avatarUrl":"/avatars/d0587bc1916029648fd5ea9a758339ea.svg","isPro":false,"fullname":"mo","user":"momo501","type":"user"},{"_id":"68d6a41703779aeebf214ccf","avatarUrl":"/avatars/4eaa149df6da50077c585dfc107c441e.svg","isPro":false,"fullname":"yong","user":"an402","type":"user"},{"_id":"68d666fdf73c8e632e733b30","avatarUrl":"/avatars/85939e3373aef74492e309691774fc6c.svg","isPro":false,"fullname":"Chuyu Qiang","user":"SHNhy","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.26730.md","query":{}}">
Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training
Published on Aug 27
· Submitted by lty on Sep 4 Abstract
Boundary-Calibrated Intervention Transfer selectively reuses past training evidence by checking contextual applicability and running bounded trials, reducing harmful updates and improving final model quality in autonomous post-training.
Large language models offer broad capabilities, but adapting them to evolving domains, tools, and requirements often entails repeated post-training. Autonomous systems automate parts of this process by proposing updates, training candidates, and using evaluation feedback to select subsequent proposals. As evidence accumulates, a central problem emerges: which past update evidence remains actionable after subsequent training has changed the parent model? An update's effect depends on its parent, data, and training stage. Treating past success as context-free permission can waste compute. If the resulting child is promoted, it can also degrade the subsequent training trajectory. We formulate this problem as conditional experience transfer and introduce Boundary-Calibrated Intervention Transfer (BCIT), a method that authorizes experience reuse before weight-changing training. BCIT binds an observed effect to its source context, checks applicability conditions, vetoes candidates with named hard conflicts, and obtains current-state evidence through a bounded training trial when needed. Fully trained candidates still face a shared adoption rule, and only observed events extend memory. On one 4B model adapted across finance reasoning, text-to-SQL, and function calling, candidate updates exhibit heterogeneous target and retention effects across the evaluated contexts. Under matched candidates, evidence, and compute, BCIT authorizes fewer harmful updates and attains higher equal-budget final-model quality than the evaluated alternatives. These results support treating experience authorization as a distinct problem in autonomous post-training.
Community
We are excited to share Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training.
Past post-training successes do not necessarily remain valid after the parent model, data, or training stage changes. We formulate this challenge as conditional experience transfer and introduce Boundary-Calibrated Intervention Transfer (BCIT), a transparent Reject / Validate / Train policy that binds update evidence to its source context, vetoes hard conflicts, and runs bounded current-parent trials when applicability is uncertain.
Across finance reasoning, text-to-SQL, and function calling adaptation with Qwen3-4B, BCIT achieves the highest equal-budget final-model quality among the evaluated shared-model alternatives. It improves the mean task score by 2.63 points over Flat-Additive (95% CI [2.10, 3.16]) while satisfying the prespecified retention margins.
📄 Paper: https://arxiv.org/abs/2608.26730
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.26730 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.26730 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.26730 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.