{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:2QE6M6KHAOIPRHZWANQJSVLOZR","short_pith_number":"pith:2QE6M6KH","schema_version":"1.0","canonical_sha256":"d409e679470390f89f36036099556ecc5f6b72f6c63d0ddf0907eb3bcb265a29","source":{"kind":"arxiv","id":"2605.09253","version":2},"attestation_state":"computed","paper":{"title":"Cornerstones or Stumbling Blocks? Deciphering the Rock Tokens in On-Policy Distillation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"High-loss Rock Tokens in on-policy distillation resist training yet add almost nothing to reasoning performance.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Dawei Li, Runchao Li, Shubhashis Roy Dipta, Yuxuan Jiang, Zhao Yang","submitted_at":"2026-05-10T01:41:43Z","abstract_excerpt":"While recent work in Reinforcement Learning with Verifiable Rewards (RLVR) has shown that a small subset of critical tokens disproportionately drives reasoning gains, an analogous token-level understanding of On-Policy Distillation (OPD) remains largely unexplored. In this work, we investigate high-loss tokens, a token type that--as the most direct signal of student-teacher mismatch under OPD's per-token KL objective--should progressively diminish as training converges according to existing studies; however, our empirical analysis shows otherwise. Even after OPD training reaches apparent satur"},"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":"2605.09253","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-10T01:41:43Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"f0fe8d86a87a7b9d7c45fa7b0042d6ff0e5c4180c3fceaf30d14b62f7deef40f","abstract_canon_sha256":"92339fc440ab4d2fbc71c519d06d3f52f4852462f92dd4ccd22dfb127e656fdb"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-02T01:03:48.286723Z","signature_b64":"aHoR6Q0KHR9wGiy+1n3CJC8ML04Ta6Izdoqy0UYSqlnaCs87GoXWMCnwy8zJbCHH3GKt28eQ/sTr2eK863KFCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d409e679470390f89f36036099556ecc5f6b72f6c63d0ddf0907eb3bcb265a29","last_reissued_at":"2026-06-02T01:03:48.286268Z","signature_status":"signed_v1","first_computed_at":"2026-06-02T01:03:48.286268Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Cornerstones or Stumbling Blocks? Deciphering the Rock Tokens in On-Policy Distillation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"High-loss Rock Tokens in on-policy distillation resist training yet add almost nothing to reasoning performance.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Dawei Li, Runchao Li, Shubhashis Roy Dipta, Yuxuan Jiang, Zhao Yang","submitted_at":"2026-05-10T01:41:43Z","abstract_excerpt":"While recent work in Reinforcement Learning with Verifiable Rewards (RLVR) has shown that a small subset of critical tokens disproportionately drives reasoning gains, an analogous token-level understanding of On-Policy Distillation (OPD) remains largely unexplored. In this work, we investigate high-loss tokens, a token type that--as the most direct signal of student-teacher mismatch under OPD's per-token KL objective--should progressively diminish as training converges according to existing studies; however, our empirical analysis shows otherwise. Even after OPD training reaches apparent satur"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"strategically bypassing these ``stumbling blocks'' can significantly streamline the alignment process, challenging the necessity of uniform token weighting and offering a more efficient paradigm for large-scale model distillation.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The causal interventions used to measure negligible functional contribution to reasoning performance are valid and complete, and that high-loss tokens identified as Rock Tokens truly have no downstream effect on model outputs.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Persistent 'Rock Tokens' in on-policy distillation resist teacher corrections, consume large gradient norms, yet add negligible value to reasoning, allowing targeted bypassing to streamline alignment.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"High-loss Rock Tokens in on-policy distillation resist training yet add almost nothing to reasoning performance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3cd6937433c316d9573fe2a6908f813c086f6d98025b071b8a2ffe296cb237b3"},"source":{"id":"2605.09253","kind":"arxiv","version":2},"verdict":{"id":"6fac69b6-7c6a-4e25-97bd-69f1b378a820","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T02:28:31.072089Z","strongest_claim":"strategically bypassing these ``stumbling blocks'' can significantly streamline the alignment process, challenging the necessity of uniform token weighting and offering a more efficient paradigm for large-scale model distillation.","one_line_summary":"Persistent 'Rock Tokens' in on-policy distillation resist teacher corrections, consume large gradient norms, yet add negligible value to reasoning, allowing targeted bypassing to streamline alignment.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The causal interventions used to measure negligible functional contribution to reasoning performance are valid and complete, and that high-loss tokens identified as Rock Tokens truly have no downstream effect on model outputs.","pith_extraction_headline":"High-loss Rock Tokens in on-policy distillation resist training yet add almost nothing to reasoning performance."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.09253/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T08:02:09.161304Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T20:34:42.788462Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T13:31:17.954079Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T10:22:49.473097Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"66825782fb90cc711e333e9db65763a0f205c05bf41fbf2f753261da0689952c"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"2509bd32ac75aa3c878a3682c0a3aef683f04124243d8863a4849671392be29f"},"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":"2605.09253","created_at":"2026-06-02T01:03:48.286327+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.09253v2","created_at":"2026-06-02T01:03:48.286327+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.09253","created_at":"2026-06-02T01:03:48.286327+00:00"},{"alias_kind":"pith_short_12","alias_value":"2QE6M6KHAOIP","created_at":"2026-06-02T01:03:48.286327+00:00"},{"alias_kind":"pith_short_16","alias_value":"2QE6M6KHAOIPRHZW","created_at":"2026-06-02T01:03:48.286327+00:00"},{"alias_kind":"pith_short_8","alias_value":"2QE6M6KH","created_at":"2026-06-02T01:03:48.286327+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2606.00448","citing_title":"When Safe Skills Collide: Measuring Compositional Risk in Agent Skill Ecosystems","ref_index":49,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/2QE6M6KHAOIPRHZWANQJSVLOZR","json":"https://pith.science/pith/2QE6M6KHAOIPRHZWANQJSVLOZR.json","graph_json":"https://pith.science/api/pith-number/2QE6M6KHAOIPRHZWANQJSVLOZR/graph.json","events_json":"https://pith.science/api/pith-number/2QE6M6KHAOIPRHZWANQJSVLOZR/events.json","paper":"https://pith.science/paper/2QE6M6KH"},"agent_actions":{"view_html":"https://pith.science/pith/2QE6M6KHAOIPRHZWANQJSVLOZR","download_json":"https://pith.science/pith/2QE6M6KHAOIPRHZWANQJSVLOZR.json","view_paper":"https://pith.science/paper/2QE6M6KH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.09253&json=true","fetch_graph":"https://pith.science/api/pith-number/2QE6M6KHAOIPRHZWANQJSVLOZR/graph.json","fetch_events":"https://pith.science/api/pith-number/2QE6M6KHAOIPRHZWANQJSVLOZR/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/2QE6M6KHAOIPRHZWANQJSVLOZR/action/timestamp_anchor","attest_storage":"https://pith.science/pith/2QE6M6KHAOIPRHZWANQJSVLOZR/action/storage_attestation","attest_author":"https://pith.science/pith/2QE6M6KHAOIPRHZWANQJSVLOZR/action/author_attestation","sign_citation":"https://pith.science/pith/2QE6M6KHAOIPRHZWANQJSVLOZR/action/citation_signature","submit_replication":"https://pith.science/pith/2QE6M6KHAOIPRHZWANQJSVLOZR/action/replication_record"}},"created_at":"2026-06-02T01:03:48.286327+00:00","updated_at":"2026-06-02T01:03:48.286327+00:00"}