{"paper":{"title":"Teacher-Guided Policy Optimization for LLM Distillation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Teacher-Guided Policy Optimization fixes uninformative feedback in reverse KL by conditioning teacher predictions on student rollouts.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Bei Li, Chunyang Xiao, Jiahao Liu, Jingang Wang, Jingbo Zhu, Junhao Ruan, Kechen Jiao, Qifan Wang, Runsong Zhao, Tong Xiao, Xin Chen, Xinyu Liu","submitted_at":"2026-05-13T09:20:03Z","abstract_excerpt":"The convergence of reinforcement learning and imitation learning has positioned Reverse KL (RKL) as a promising paradigm for on-policy LLM distillation, aiming to unify exploration with teacher supervision. However, we identify a critical limitation: when the student and teacher distributions diverge significantly, standard RKL often fails to yield meaningful improvement due to uninformative negative feedback. To address this inefficiency, we propose Teacher-Guided Policy Optimization (TGPO), an on-policy algorithm that incorporates dense directional guidance by leveraging teacher predictions "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"TGPO significantly outperforms standard baselines and is robust to different teachers on complex reasoning benchmarks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That conditioning teacher predictions on the student's rollout will reliably produce informative directional guidance even when student and teacher distributions diverge substantially.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"TGPO improves on-policy LLM distillation by using teacher predictions conditioned on student rollouts to supply informative guidance when the two distributions diverge.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Teacher-Guided Policy Optimization fixes uninformative feedback in reverse KL by conditioning teacher predictions on student rollouts.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"8a2901be0c3cc032ae152b0d89721a4915a10996bb763507ee2b68437eb85823"},"source":{"id":"2605.13230","kind":"arxiv","version":1},"verdict":{"id":"34e18f50-4e1d-47e3-b594-ffb42a21d78a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:08:08.482193Z","strongest_claim":"TGPO significantly outperforms standard baselines and is robust to different teachers on complex reasoning benchmarks.","one_line_summary":"TGPO improves on-policy LLM distillation by using teacher predictions conditioned on student rollouts to supply informative guidance when the two distributions diverge.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That conditioning teacher predictions on the student's rollout will reliably produce informative directional guidance even when student and teacher distributions diverge substantially.","pith_extraction_headline":"Teacher-Guided Policy Optimization fixes uninformative feedback in reverse KL by conditioning teacher predictions on student rollouts."},"references":{"count":23,"sample":[{"doi":"","year":null,"title":"SFT or RL? An Early Investigation into Training R1-Like Reasoning Large Vision-Language Models","work_id":"a521360c-8673-4d0d-a3a3-6eb9f7a71b90","ref_index":1,"cited_arxiv_id":"2504.11468","is_internal_anchor":true},{"doi":"","year":null,"title":"SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training","work_id":"258dd934-025c-47f5-b4f6-5a0c1c338cc6","ref_index":2,"cited_arxiv_id":"2501.17161","is_internal_anchor":true},{"doi":"","year":null,"title":"Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge","work_id":"28ea1282-d657-4c61-a83c-f1249be6d6b1","ref_index":3,"cited_arxiv_id":"1803.05457","is_internal_anchor":true},{"doi":"","year":null,"title":"Process Reinforcement through Implicit Rewards","work_id":"c31a2126-86f9-44f3-91f3-208d0fc1463a","ref_index":4,"cited_arxiv_id":"2502.01456","is_internal_anchor":true},{"doi":"","year":null,"title":"MiniLLM: On-Policy Distillation of Large Language Models","work_id":"16edb291-dd18-41c5-8486-c6c715ec5311","ref_index":5,"cited_arxiv_id":"2306.08543","is_internal_anchor":true}],"resolved_work":23,"snapshot_sha256":"6b4db49c37be5a137b76f46ca181a3ed485126ffe3b526aef6223a9f420410de","internal_anchors":18},"formal_canon":{"evidence_count":2,"snapshot_sha256":"38b4428d457f6357cc5dd19881c66c582aecf2de28b6117742a1e70fd9a027cd"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}