{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:VPZTCO4J4EG4HSUEEIFR2J7HZT","merge_version":"pith-open-graph-merge-v1","event_count":3,"valid_event_count":3,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"2c4910f7ce364db0a93c72122ee6e645a59abedf6cfd9fab5f920fd384541c66","cross_cats_sorted":["cs.AI"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T09:20:03Z","title_canon_sha256":"0e90ea93b8cdaae16215c07b5d8dfcd34c3814bd0bf4212ce3a2310bad428029"},"schema_version":"1.0","source":{"id":"2605.13230","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.13230","created_at":"2026-05-18T02:44:49Z"},{"alias_kind":"arxiv_version","alias_value":"2605.13230v1","created_at":"2026-05-18T02:44:49Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13230","created_at":"2026-05-18T02:44:49Z"},{"alias_kind":"pith_short_12","alias_value":"VPZTCO4J4EG4","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"VPZTCO4J4EG4HSUE","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"VPZTCO4J","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:0b9974633aa55ba8e9ec67d3753b5ae565de6f4e615976a6285b195d5642d735","target":"graph","created_at":"2026-05-18T02:44:49Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"TGPO significantly outperforms standard baselines and is robust to different teachers on complex reasoning benchmarks."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That conditioning teacher predictions on the student's rollout will reliably produce informative directional guidance even when student and teacher distributions diverge substantially."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"TGPO improves on-policy LLM distillation by using teacher predictions conditioned on student rollouts to supply informative guidance when the two distributions diverge."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Teacher-Guided Policy Optimization fixes uninformative feedback in reverse KL by conditioning teacher predictions on student rollouts."}],"snapshot_sha256":"8a2901be0c3cc032ae152b0d89721a4915a10996bb763507ee2b68437eb85823"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"38b4428d457f6357cc5dd19881c66c582aecf2de28b6117742a1e70fd9a027cd"},"paper":{"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 ","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","cross_cats":["cs.AI"],"headline":"Teacher-Guided Policy Optimization fixes uninformative feedback in reverse KL by conditioning teacher predictions on student rollouts.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T09:20:03Z","title":"Teacher-Guided Policy Optimization for LLM Distillation"},"references":{"count":23,"internal_anchors":18,"resolved_work":23,"sample":[{"cited_arxiv_id":"2504.11468","doi":"","is_internal_anchor":true,"ref_index":1,"title":"SFT or RL? An Early Investigation into Training R1-Like Reasoning Large Vision-Language Models","work_id":"a521360c-8673-4d0d-a3a3-6eb9f7a71b90","year":null},{"cited_arxiv_id":"2501.17161","doi":"","is_internal_anchor":true,"ref_index":2,"title":"SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training","work_id":"258dd934-025c-47f5-b4f6-5a0c1c338cc6","year":null},{"cited_arxiv_id":"1803.05457","doi":"","is_internal_anchor":true,"ref_index":3,"title":"Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge","work_id":"28ea1282-d657-4c61-a83c-f1249be6d6b1","year":null},{"cited_arxiv_id":"2502.01456","doi":"","is_internal_anchor":true,"ref_index":4,"title":"Process Reinforcement through Implicit Rewards","work_id":"c31a2126-86f9-44f3-91f3-208d0fc1463a","year":null},{"cited_arxiv_id":"2306.08543","doi":"","is_internal_anchor":true,"ref_index":5,"title":"MiniLLM: On-Policy Distillation of Large Language Models","work_id":"16edb291-dd18-41c5-8486-c6c715ec5311","year":null}],"snapshot_sha256":"6b4db49c37be5a137b76f46ca181a3ed485126ffe3b526aef6223a9f420410de"},"source":{"id":"2605.13230","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-14T20:08:08.482193Z","id":"34e18f50-4e1d-47e3-b594-ffb42a21d78a","model_set":{"reader":"grok-4.3"},"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","pith_extraction_headline":"Teacher-Guided Policy Optimization fixes uninformative feedback in reverse KL by conditioning teacher predictions on student rollouts.","strongest_claim":"TGPO significantly outperforms standard baselines and is robust to different teachers on complex reasoning benchmarks.","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."}},"verdict_id":"34e18f50-4e1d-47e3-b594-ffb42a21d78a"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:ccd07853bbe0b77eb80d4b1b0bbedd163ede28c5866ef36a0c6dd85c95402cb9","target":"record","created_at":"2026-05-18T02:44:49Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"2c4910f7ce364db0a93c72122ee6e645a59abedf6cfd9fab5f920fd384541c66","cross_cats_sorted":["cs.AI"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T09:20:03Z","title_canon_sha256":"0e90ea93b8cdaae16215c07b5d8dfcd34c3814bd0bf4212ce3a2310bad428029"},"schema_version":"1.0","source":{"id":"2605.13230","kind":"arxiv","version":1}},"canonical_sha256":"abf3313b89e10dc3ca84220b1d27e7ccd7906413524cf5e18781579e1ba986ce","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"abf3313b89e10dc3ca84220b1d27e7ccd7906413524cf5e18781579e1ba986ce","first_computed_at":"2026-05-18T02:44:49.594387Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T02:44:49.594387Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"NcPYvEzGSECok955zBjajnBefCoNVmnqeNdSnEfzbDUo16mMRPvD7aQ0j/rL4QE5aA7yPnj9AHAdJZXCP3QVCw==","signature_status":"signed_v1","signed_at":"2026-05-18T02:44:49.594892Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.13230","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:ccd07853bbe0b77eb80d4b1b0bbedd163ede28c5866ef36a0c6dd85c95402cb9","sha256:0b9974633aa55ba8e9ec67d3753b5ae565de6f4e615976a6285b195d5642d735","sha256:2086d57386abb1959236c3ec5ccc091633c7c961ecb773b503430b980a6f2f30"],"state_sha256":"bf306cb1145574e736aa7402e9feb0abd7ff0f6a130fcac9195e7b7d6a8e394f"}