{"paper":{"title":"TIP: Token Importance in On-Policy Distillation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Informative tokens in on-policy distillation come from high student entropy positions and low-entropy positions with high teacher divergence where the student is overconfident and wrong.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Alborz Geramifard, Hejian Sang, Ran He, Yuanda Xu, Zhengze Zhou, Zhipeng Wang","submitted_at":"2026-04-15T16:58:24Z","abstract_excerpt":"On-policy knowledge distillation (OPD) trains a student on its own rollouts under token-level supervision from a teacher. Not all token positions matter equally, but existing views of token importance are incomplete. We ask a direct question: which tokens carry the most useful learning signal in OPD? Our answer is that informative tokens come from two regions: positions with high student entropy, and positions with low student entropy plus high teacher--student divergence, where the student is overconfident and wrong.\n  Empirically, student entropy is a strong first-order proxy: retaining $50\\"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"informative tokens come from two regions: positions with high student entropy, and positions with low student entropy plus high teacher--student divergence, where the student is overconfident and wrong.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That teacher-student divergence reliably flags cases where the student is factually wrong rather than simply differing in style or valid alternative answers.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"In on-policy distillation, tokens with high student entropy or low entropy plus high teacher divergence provide dense corrective signal, allowing effective training on under 20% of tokens across math and planning tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Informative tokens in on-policy distillation come from high student entropy positions and low-entropy positions with high teacher divergence where the student is overconfident and wrong.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d54ce8333edf6342f15f10dae62c2e9c86bcd444aed6d25145f84fa7b523d04b"},"source":{"id":"2604.14084","kind":"arxiv","version":3},"verdict":{"id":"14de82d4-9b9c-440d-a81f-074f75090484","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T13:23:43.447420Z","strongest_claim":"informative tokens come from two regions: positions with high student entropy, and positions with low student entropy plus high teacher--student divergence, where the student is overconfident and wrong.","one_line_summary":"In on-policy distillation, tokens with high student entropy or low entropy plus high teacher divergence provide dense corrective signal, allowing effective training on under 20% of tokens across math and planning tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That teacher-student divergence reliably flags cases where the student is factually wrong rather than simply differing in style or valid alternative answers.","pith_extraction_headline":"Informative tokens in on-policy distillation come from high student entropy positions and low-entropy positions with high teacher divergence where the student is overconfident and wrong."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.14084/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}