{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:WJBPH6XG6VAQD7ULFG3MQVQXHC","short_pith_number":"pith:WJBPH6XG","schema_version":"1.0","canonical_sha256":"b242f3fae6f54101fe8b29b6c85617388dcc106a4f81952f7980221ced6926a8","source":{"kind":"arxiv","id":"2510.06062","version":2},"attestation_state":"computed","paper":{"title":"When Importance Sampling Misallocates Credit: Asymmetric Ratios for Outcome-Supervised RL","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Fuzheng Zhang, Guorui Zhou, Jiakang Wang, Kun Gai, Lei Lin, Ling Pan, Qingpeng Cai, Runze Liu, Wenping Hu, Xiu Li","submitted_at":"2025-10-07T15:54:24Z","abstract_excerpt":"Reinforcement learning (RL) has shown great promise in large language models (LLMs) post-training, which typically rely on token-level clipping to maintain stability during optimization. Despite the empirical success of GRPO-style methods, we identify a fundamental and previously overlooked challenge in this popular Outcome-Supervised RL (OSRL) paradigm. We reveal that in OSRL, where advantages are shared across tokens within a response, importance sampling (IS) ratios deviate from their traditional purpose of distribution correction as in classic RL, which become token-level weights that allo"},"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":false},"canonical_record":{"source":{"id":"2510.06062","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-10-07T15:54:24Z","cross_cats_sorted":[],"title_canon_sha256":"6fabb6417347564004a63aec21336baa57e57ec92199b3164180bdd9582977bf","abstract_canon_sha256":"f04a7db0ce9a8aaca46bb6cb8c9ce60a53c7335b72a9ccfd89fb80ac0af6e6b5"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:00:25.447550Z","signature_b64":"009DKiyTCpNcuriqVeM8BTUkKj/yn/unUrl3N97RU5lZUVSqikY6gntZ9ckl2vVA81Febmb8t7195WorgdiQBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b242f3fae6f54101fe8b29b6c85617388dcc106a4f81952f7980221ced6926a8","last_reissued_at":"2026-05-20T00:00:25.446716Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:00:25.446716Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"When Importance Sampling Misallocates Credit: Asymmetric Ratios for Outcome-Supervised RL","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Fuzheng Zhang, Guorui Zhou, Jiakang Wang, Kun Gai, Lei Lin, Ling Pan, Qingpeng Cai, Runze Liu, Wenping Hu, Xiu Li","submitted_at":"2025-10-07T15:54:24Z","abstract_excerpt":"Reinforcement learning (RL) has shown great promise in large language models (LLMs) post-training, which typically rely on token-level clipping to maintain stability during optimization. Despite the empirical success of GRPO-style methods, we identify a fundamental and previously overlooked challenge in this popular Outcome-Supervised RL (OSRL) paradigm. We reveal that in OSRL, where advantages are shared across tokens within a response, importance sampling (IS) ratios deviate from their traditional purpose of distribution correction as in classic RL, which become token-level weights that allo"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2510.06062","kind":"arxiv","version":2},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2510.06062/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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2510.06062","created_at":"2026-05-20T00:00:25.446867+00:00"},{"alias_kind":"arxiv_version","alias_value":"2510.06062v2","created_at":"2026-05-20T00:00:25.446867+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2510.06062","created_at":"2026-05-20T00:00:25.446867+00:00"},{"alias_kind":"pith_short_12","alias_value":"WJBPH6XG6VAQ","created_at":"2026-05-20T00:00:25.446867+00:00"},{"alias_kind":"pith_short_16","alias_value":"WJBPH6XG6VAQD7UL","created_at":"2026-05-20T00:00:25.446867+00:00"},{"alias_kind":"pith_short_8","alias_value":"WJBPH6XG","created_at":"2026-05-20T00:00:25.446867+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":10,"internal_anchor_count":10,"sample":[{"citing_arxiv_id":"2605.22703","citing_title":"Clipping Bottleneck: Stabilizing RLVR via Stochastic Recovery of Near-Boundary Signals","ref_index":20,"is_internal_anchor":true},{"citing_arxiv_id":"2605.20865","citing_title":"Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19425","citing_title":"When to Stop Reusing: Dynamic Gradient Gating for Sample-Efficient RLVR","ref_index":31,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19282","citing_title":"Rethinking Muon Beyond Pretraining: Spectral Failures and High-Pass Remedies for VLA and RLVR","ref_index":34,"is_internal_anchor":true},{"citing_arxiv_id":"2602.15620","citing_title":"STAPO: Stabilizing Reinforcement Learning for LLMs by Silencing Rare Spurious Tokens","ref_index":28,"is_internal_anchor":true},{"citing_arxiv_id":"2605.11922","citing_title":"StepCodeReasoner: Aligning Code Reasoning with Stepwise Execution Traces via Reinforcement Learning","ref_index":75,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08737","citing_title":"The Extrapolation Cliff in On-Policy Distillation of Near-Deterministic Structured Outputs","ref_index":39,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18530","citing_title":"OGER: A Robust Offline-Guided Exploration Reward for Hybrid Reinforcement Learning","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2605.04077","citing_title":"Balanced Aggregation: Understanding and Fixing Aggregation Bias in GRPO","ref_index":19,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18578","citing_title":"Bounded Ratio Reinforcement Learning","ref_index":30,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/WJBPH6XG6VAQD7ULFG3MQVQXHC","json":"https://pith.science/pith/WJBPH6XG6VAQD7ULFG3MQVQXHC.json","graph_json":"https://pith.science/api/pith-number/WJBPH6XG6VAQD7ULFG3MQVQXHC/graph.json","events_json":"https://pith.science/api/pith-number/WJBPH6XG6VAQD7ULFG3MQVQXHC/events.json","paper":"https://pith.science/paper/WJBPH6XG"},"agent_actions":{"view_html":"https://pith.science/pith/WJBPH6XG6VAQD7ULFG3MQVQXHC","download_json":"https://pith.science/pith/WJBPH6XG6VAQD7ULFG3MQVQXHC.json","view_paper":"https://pith.science/paper/WJBPH6XG","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2510.06062&json=true","fetch_graph":"https://pith.science/api/pith-number/WJBPH6XG6VAQD7ULFG3MQVQXHC/graph.json","fetch_events":"https://pith.science/api/pith-number/WJBPH6XG6VAQD7ULFG3MQVQXHC/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/WJBPH6XG6VAQD7ULFG3MQVQXHC/action/timestamp_anchor","attest_storage":"https://pith.science/pith/WJBPH6XG6VAQD7ULFG3MQVQXHC/action/storage_attestation","attest_author":"https://pith.science/pith/WJBPH6XG6VAQD7ULFG3MQVQXHC/action/author_attestation","sign_citation":"https://pith.science/pith/WJBPH6XG6VAQD7ULFG3MQVQXHC/action/citation_signature","submit_replication":"https://pith.science/pith/WJBPH6XG6VAQD7ULFG3MQVQXHC/action/replication_record"}},"created_at":"2026-05-20T00:00:25.446867+00:00","updated_at":"2026-05-20T00:00:25.446867+00:00"}