{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:VFLFREZPN6POFUG3NPVHGIOBY4","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"de0bcebca2a87255dcfef7f5533ad79685999585cf31a57ff359212454eb5959","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-11T19:44:59Z","title_canon_sha256":"b248cb51e667023b57534f1c43b260f0bb7b7d2661aab3993870fca78c420d80"},"schema_version":"1.0","source":{"id":"2605.11182","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.11182","created_at":"2026-05-26T01:03:33Z"},{"alias_kind":"arxiv_version","alias_value":"2605.11182v2","created_at":"2026-05-26T01:03:33Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.11182","created_at":"2026-05-26T01:03:33Z"},{"alias_kind":"pith_short_12","alias_value":"VFLFREZPN6PO","created_at":"2026-05-26T01:03:33Z"},{"alias_kind":"pith_short_16","alias_value":"VFLFREZPN6POFUG3","created_at":"2026-05-26T01:03:33Z"},{"alias_kind":"pith_short_8","alias_value":"VFLFREZP","created_at":"2026-05-26T01:03:33Z"}],"graph_snapshots":[{"event_id":"sha256:2ffdac2896e3e9d62710e13d0604c789fb6330719f9667fa9f4087ebc23b6da1","target":"graph","created_at":"2026-05-26T01:03:33Z","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":"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."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"On-policy distillation fails in LLMs due to distribution mismatch, biased gradients, and privileged information aggregation but targeted fixes restore effectiveness."}],"snapshot_sha256":"832de19b3cd4f2b63119cd12226e252ddd28bde55f16c215f062eb805aecbf04"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[{"findings_count":0,"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"}],"endpoint":"/pith/2605.11182/integrity.json","findings":[],"snapshot_sha256":"e3f86565937dc8f02f8385eeaa5b3f6abf8d1dad995e7e7b0150e5577541d36e","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"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","authors_text":"Ge Liu, Hongyu Lu, Siqi Zhu, Weiye Shi, Xuyan Ye","cross_cats":[],"headline":"On-policy distillation fails in LLMs due to distribution mismatch, biased gradients, and privileged information aggregation but targeted fixes restore effectiveness.","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-11T19:44:59Z","title":"The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.11182","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-13T02:13:55.930380Z","id":"25146744-6cee-46de-be8c-ce7581fd0189","model_set":{"reader":"grok-4.3"},"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","pith_extraction_headline":"On-policy distillation fails in LLMs due to distribution mismatch, biased gradients, and privileged information aggregation but targeted fixes restore effectiveness.","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.","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."}},"verdict_id":"25146744-6cee-46de-be8c-ce7581fd0189"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:8c24232f1936cd0a61a688ce594a1b557c6279421730d638eb6f913249bcb7a5","target":"record","created_at":"2026-05-26T01:03:33Z","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":"de0bcebca2a87255dcfef7f5533ad79685999585cf31a57ff359212454eb5959","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-11T19:44:59Z","title_canon_sha256":"b248cb51e667023b57534f1c43b260f0bb7b7d2661aab3993870fca78c420d80"},"schema_version":"1.0","source":{"id":"2605.11182","kind":"arxiv","version":2}},"canonical_sha256":"a95658932f6f9ee2d0db6bea7321c1c71826e3bb411c55a1db19227701c45158","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"a95658932f6f9ee2d0db6bea7321c1c71826e3bb411c55a1db19227701c45158","first_computed_at":"2026-05-26T01:03:33.409214Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-26T01:03:33.409214Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"7trjWTxlCnNDCdEdnrPq1XMKZyi9NBY/HqfKqfQxdFIPIu/zbAEYKbjkqJtSf44W119BDXpUGBSB3BSLG5XqCg==","signature_status":"signed_v1","signed_at":"2026-05-26T01:03:33.409959Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.11182","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:8c24232f1936cd0a61a688ce594a1b557c6279421730d638eb6f913249bcb7a5","sha256:2ffdac2896e3e9d62710e13d0604c789fb6330719f9667fa9f4087ebc23b6da1"],"state_sha256":"24caa778e825bddd13579e9f4e73058ed521f74c4666bd5c6103c25ce8d2d6ed"}