{"paper":{"title":"The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"On-policy distillation fails in LLMs due to distribution mismatch, biased gradients, and privileged information aggregation but targeted fixes restore effectiveness.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Ge Liu, Hongyu Lu, Siqi Zhu, Weiye Shi, Xuyan Ye","submitted_at":"2026-05-11T19:44:59Z","abstract_excerpt":"On-policy distillation (OPD) and on-policy self-distillation (OPSD) have emerged as promising post-training methods for large language models, offering dense token-level supervision on trajectories sampled from the model's own policy. However, existing results on their effectiveness remain mixed: while OP(S)D has shown promise in system prompt and knowledge internalization, recent studies also report instability and degradation. In this work, we present a comprehensive empirical study of when OPD and OPSD work, when they fail, and why. We find that OPD on mathematical reasoning is highly sensi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We identify three failure mechanisms: (1) distribution mismatch between teacher and student caused by conditioning on student-generated prefixes, (2) optimization instability from biased TopK reverse-KL gradients, and (3) an OPSD-specific limitation where the student learns a PI-free policy that aggregates PI-conditioned teachers, which is insufficient when PI is instance-specific. We further show that stop-gradient TopK objectives, RLVR-adapted teachers, and SFT-stabilized students mitigate these failures.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The tested settings (mathematical reasoning trajectories and system-prompt/alignment PI) are representative enough that the three failure mechanisms and proposed fixes will apply to other LLM tasks, model scales, and data distributions without additional confounding factors.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"On-policy distillation for LLMs is sensitive to teacher choice and loss design, while self-distillation fails on instance-specific information but succeeds on shared rules, with stop-gradient TopK, adapted teachers, and SFT stabilization as mitigations.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"On-policy distillation fails in LLMs due to distribution mismatch, biased gradients, and privileged information aggregation but targeted fixes restore effectiveness.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"832de19b3cd4f2b63119cd12226e252ddd28bde55f16c215f062eb805aecbf04"},"source":{"id":"2605.11182","kind":"arxiv","version":2},"verdict":{"id":"25146744-6cee-46de-be8c-ce7581fd0189","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T02:13:55.930380Z","strongest_claim":"We identify three failure mechanisms: (1) distribution mismatch between teacher and student caused by conditioning on student-generated prefixes, (2) optimization instability from biased TopK reverse-KL gradients, and (3) an OPSD-specific limitation where the student learns a PI-free policy that aggregates PI-conditioned teachers, which is insufficient when PI is instance-specific. We further show that stop-gradient TopK objectives, RLVR-adapted teachers, and SFT-stabilized students mitigate these failures.","one_line_summary":"On-policy distillation for LLMs is sensitive to teacher choice and loss design, while self-distillation fails on instance-specific information but succeeds on shared rules, with stop-gradient TopK, adapted teachers, and SFT stabilization as mitigations.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The tested settings (mathematical reasoning trajectories and system-prompt/alignment PI) are representative enough that the three failure mechanisms and proposed fixes will apply to other LLM tasks, model scales, and data distributions without additional confounding factors.","pith_extraction_headline":"On-policy distillation fails in LLMs due to distribution mismatch, biased gradients, and privileged information aggregation but targeted fixes restore effectiveness."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.11182/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T04:42:00.971795Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T13:33:52.618077Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T10:01:17.612652Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T08:42:08.036066Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"e3f86565937dc8f02f8385eeaa5b3f6abf8d1dad995e7e7b0150e5577541d36e"},"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"}