{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:IBATJDLOU3LETEJXTQF67NBOJN","short_pith_number":"pith:IBATJDLO","schema_version":"1.0","canonical_sha256":"4041348d6ea6d64991379c0befb42e4b55be755766871ddb865528ac4527a08d","source":{"kind":"arxiv","id":"2605.07804","version":2},"attestation_state":"computed","paper":{"title":"Prune-OPD: Efficient and Reliable On-Policy Distillation for Long-Horizon Reasoning","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Prune-OPD makes on-policy distillation for long-horizon reasoning more efficient by pruning unreliable teacher rewards in real time.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Jing Tang, Minrui Xu, Xiaodan Liang, Yifan Song, Yiwei Wang, Yongxin Wang, Zhicheng Yang, Zhijiang Guo","submitted_at":"2026-05-08T14:38:53Z","abstract_excerpt":"On-policy distillation (OPD) leverages dense teacher rewards to enhance reasoning models. However, scaling OPD to long-horizon tasks exposes a critical flaw: as the student's generated prefix inevitably diverges from the teacher's thought process, the teacher's dense reward loses local exploitability. Continuing to generate and evaluate tokens on these ``drifted'' trajectories not only degrades reward quality but also incurs massive computational waste. To address this, we introduce \\textbf{Prune-OPD}, a framework that dynamically aligns training budgets with supervision quality. By continuous"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":true},"canonical_record":{"source":{"id":"2605.07804","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-08T14:38:53Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"1ddfef6338d71579cc32c9d572291023565d114402883f18cd7f0ae117a97825","abstract_canon_sha256":"6e600b008c47516456ddd1fa78f0bb9609bf7056c0cd27b6bb77a80754a25007"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-29T02:05:46.259561Z","signature_b64":"t5wBwYhB6tV4kwQZ8etzSIx2qVGyOonUmZUOMKCCg20Ca2qyzTRbgfsgyWFp/xaq6iGP9Zu54CBk9xho8FsVBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4041348d6ea6d64991379c0befb42e4b55be755766871ddb865528ac4527a08d","last_reissued_at":"2026-05-29T02:05:46.258958Z","signature_status":"signed_v1","first_computed_at":"2026-05-29T02:05:46.258958Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Prune-OPD: Efficient and Reliable On-Policy Distillation for Long-Horizon Reasoning","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Prune-OPD makes on-policy distillation for long-horizon reasoning more efficient by pruning unreliable teacher rewards in real time.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Jing Tang, Minrui Xu, Xiaodan Liang, Yifan Song, Yiwei Wang, Yongxin Wang, Zhicheng Yang, Zhijiang Guo","submitted_at":"2026-05-08T14:38:53Z","abstract_excerpt":"On-policy distillation (OPD) leverages dense teacher rewards to enhance reasoning models. However, scaling OPD to long-horizon tasks exposes a critical flaw: as the student's generated prefix inevitably diverges from the teacher's thought process, the teacher's dense reward loses local exploitability. Continuing to generate and evaluate tokens on these ``drifted'' trajectories not only degrades reward quality but also incurs massive computational waste. To address this, we introduce \\textbf{Prune-OPD}, a framework that dynamically aligns training budgets with supervision quality. By continuous"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Prune-OPD reduces training time by 37.6%--68.0% while preserving, and often improving, performance on challenging benchmarks (AMC, AIME, HMMT) by dynamically aligning computation with supervision reliability across diverse teacher-student combinations.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That top-k overlap between student and teacher predictions is a reliable real-time indicator of when dense teacher rewards lose local exploitability, and that monotonic down-weighting plus truncation does not discard critical learning signals needed for long-horizon improvement.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Prune-OPD dynamically prunes unreliable teacher rewards in on-policy distillation by monitoring prefix drift via top-k overlap, reducing training time 37.6-68% on AMC/AIME/HMMT while preserving or improving performance.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Prune-OPD makes on-policy distillation for long-horizon reasoning more efficient by pruning unreliable teacher rewards in real time.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"47398ec2ba48b8a26806dd2bbf5c591a01be0727c8a7a91bc7843afc7f74178d"},"source":{"id":"2605.07804","kind":"arxiv","version":2},"verdict":{"id":"ee240024-508d-47e7-a45d-187ed1cfa5ba","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-11T03:15:11.248353Z","strongest_claim":"Prune-OPD reduces training time by 37.6%--68.0% while preserving, and often improving, performance on challenging benchmarks (AMC, AIME, HMMT) by dynamically aligning computation with supervision reliability across diverse teacher-student combinations.","one_line_summary":"Prune-OPD dynamically prunes unreliable teacher rewards in on-policy distillation by monitoring prefix drift via top-k overlap, reducing training time 37.6-68% on AMC/AIME/HMMT while preserving or improving performance.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That top-k overlap between student and teacher predictions is a reliable real-time indicator of when dense teacher rewards lose local exploitability, and that monotonic down-weighting plus truncation does not discard critical learning signals needed for long-horizon improvement.","pith_extraction_headline":"Prune-OPD makes on-policy distillation for long-horizon reasoning more efficient by pruning unreliable teacher rewards in real time."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.07804/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T10:22:02.734357Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-20T04:50:05.693007Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T15:31:18.842395Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T11:32:17.673878Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"3a3d2a13ab9d23d2f5a72728efba70f7d75dfeb91beb8e649f3c0fc591e500e5"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"417f5c0c6787c8438a67801a88bebd7d79258251d92a7fd3d4856ae90d9331d2"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.07804","created_at":"2026-05-29T02:05:46.259052+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.07804v2","created_at":"2026-05-29T02:05:46.259052+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.07804","created_at":"2026-05-29T02:05:46.259052+00:00"},{"alias_kind":"pith_short_12","alias_value":"IBATJDLOU3LE","created_at":"2026-05-29T02:05:46.259052+00:00"},{"alias_kind":"pith_short_16","alias_value":"IBATJDLOU3LETEJX","created_at":"2026-05-29T02:05:46.259052+00:00"},{"alias_kind":"pith_short_8","alias_value":"IBATJDLO","created_at":"2026-05-29T02:05:46.259052+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/IBATJDLOU3LETEJXTQF67NBOJN","json":"https://pith.science/pith/IBATJDLOU3LETEJXTQF67NBOJN.json","graph_json":"https://pith.science/api/pith-number/IBATJDLOU3LETEJXTQF67NBOJN/graph.json","events_json":"https://pith.science/api/pith-number/IBATJDLOU3LETEJXTQF67NBOJN/events.json","paper":"https://pith.science/paper/IBATJDLO"},"agent_actions":{"view_html":"https://pith.science/pith/IBATJDLOU3LETEJXTQF67NBOJN","download_json":"https://pith.science/pith/IBATJDLOU3LETEJXTQF67NBOJN.json","view_paper":"https://pith.science/paper/IBATJDLO","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.07804&json=true","fetch_graph":"https://pith.science/api/pith-number/IBATJDLOU3LETEJXTQF67NBOJN/graph.json","fetch_events":"https://pith.science/api/pith-number/IBATJDLOU3LETEJXTQF67NBOJN/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/IBATJDLOU3LETEJXTQF67NBOJN/action/timestamp_anchor","attest_storage":"https://pith.science/pith/IBATJDLOU3LETEJXTQF67NBOJN/action/storage_attestation","attest_author":"https://pith.science/pith/IBATJDLOU3LETEJXTQF67NBOJN/action/author_attestation","sign_citation":"https://pith.science/pith/IBATJDLOU3LETEJXTQF67NBOJN/action/citation_signature","submit_replication":"https://pith.science/pith/IBATJDLOU3LETEJXTQF67NBOJN/action/replication_record"}},"created_at":"2026-05-29T02:05:46.259052+00:00","updated_at":"2026-05-29T02:05:46.259052+00:00"}