{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:3JQDU2X5BXM2OCQ2IQ63GQDH3S","short_pith_number":"pith:3JQDU2X5","schema_version":"1.0","canonical_sha256":"da603a6afd0dd9a70a1a443db34067dc9d1c1b34f13373f49a96d6046b7c2b68","source":{"kind":"arxiv","id":"2510.08977","version":2},"attestation_state":"computed","paper":{"title":"Breaking the Self-Confirming Loop: Diagnosing and Mitigating Systemic Reward Bias in Self-Rewarding RL","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Boyuan Pan, Chuyi Tan, Jiayi Shi, Ji Zhang, Kan Li, Peiwen Yuan, Shaoxiong Feng, Xinglin Wang, Yao Hu, Yiwei Li, Yueqi Zhang","submitted_at":"2025-10-10T03:38:17Z","abstract_excerpt":"Reinforcement learning with verifiable rewards (RLVR) efficiently scales the reasoning ability of large language models (LLMs) but is bottlenecked by scarce labeled data. Reinforcement learning with intrinsic rewards (RLIR) offers a scalable alternative via self-rewarding, yet often suffers from instability and inferior performance. We trace this gap to a systemic bias in confidence-coupled self-rewarding: the model tends to over-reward high-confidence mistakes, forming a self-confirming loop. We quantify this feedback-loop bias with three metrics: reward noise magnitude (rho_noise), policy-re"},"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":"2510.08977","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-10-10T03:38:17Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"6fd71bac58d0c20a2db112d9f81d60c503419a619395dc081c5980598154ec73","abstract_canon_sha256":"ea44a5c1a4dd15c0dfebc002406976c960d6bc868f1ae9e9290aed62138cc3fe"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-03T01:05:45.626867Z","signature_b64":"d969Dnwt21SklBEkoXZrFQoKQyFWE1sU46HHtSypvLqV2VyimCiyGfof4V99q0h5gNWMx4JBHYnEwI8DQoEHAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"da603a6afd0dd9a70a1a443db34067dc9d1c1b34f13373f49a96d6046b7c2b68","last_reissued_at":"2026-06-03T01:05:45.626366Z","signature_status":"signed_v1","first_computed_at":"2026-06-03T01:05:45.626366Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Breaking the Self-Confirming Loop: Diagnosing and Mitigating Systemic Reward Bias in Self-Rewarding RL","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Boyuan Pan, Chuyi Tan, Jiayi Shi, Ji Zhang, Kan Li, Peiwen Yuan, Shaoxiong Feng, Xinglin Wang, Yao Hu, Yiwei Li, Yueqi Zhang","submitted_at":"2025-10-10T03:38:17Z","abstract_excerpt":"Reinforcement learning with verifiable rewards (RLVR) efficiently scales the reasoning ability of large language models (LLMs) but is bottlenecked by scarce labeled data. Reinforcement learning with intrinsic rewards (RLIR) offers a scalable alternative via self-rewarding, yet often suffers from instability and inferior performance. We trace this gap to a systemic bias in confidence-coupled self-rewarding: the model tends to over-reward high-confidence mistakes, forming a self-confirming loop. We quantify this feedback-loop bias with three metrics: reward noise magnitude (rho_noise), policy-re"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2510.08977","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/2510.08977/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":"2510.08977","created_at":"2026-06-03T01:05:45.626427+00:00"},{"alias_kind":"arxiv_version","alias_value":"2510.08977v2","created_at":"2026-06-03T01:05:45.626427+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2510.08977","created_at":"2026-06-03T01:05:45.626427+00:00"},{"alias_kind":"pith_short_12","alias_value":"3JQDU2X5BXM2","created_at":"2026-06-03T01:05:45.626427+00:00"},{"alias_kind":"pith_short_16","alias_value":"3JQDU2X5BXM2OCQ2","created_at":"2026-06-03T01:05:45.626427+00:00"},{"alias_kind":"pith_short_8","alias_value":"3JQDU2X5","created_at":"2026-06-03T01:05:45.626427+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2603.19880","citing_title":"What If Consensus Lies? Selective-Complementary Reinforcement Learning at Test Time","ref_index":16,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3JQDU2X5BXM2OCQ2IQ63GQDH3S","json":"https://pith.science/pith/3JQDU2X5BXM2OCQ2IQ63GQDH3S.json","graph_json":"https://pith.science/api/pith-number/3JQDU2X5BXM2OCQ2IQ63GQDH3S/graph.json","events_json":"https://pith.science/api/pith-number/3JQDU2X5BXM2OCQ2IQ63GQDH3S/events.json","paper":"https://pith.science/paper/3JQDU2X5"},"agent_actions":{"view_html":"https://pith.science/pith/3JQDU2X5BXM2OCQ2IQ63GQDH3S","download_json":"https://pith.science/pith/3JQDU2X5BXM2OCQ2IQ63GQDH3S.json","view_paper":"https://pith.science/paper/3JQDU2X5","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2510.08977&json=true","fetch_graph":"https://pith.science/api/pith-number/3JQDU2X5BXM2OCQ2IQ63GQDH3S/graph.json","fetch_events":"https://pith.science/api/pith-number/3JQDU2X5BXM2OCQ2IQ63GQDH3S/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3JQDU2X5BXM2OCQ2IQ63GQDH3S/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3JQDU2X5BXM2OCQ2IQ63GQDH3S/action/storage_attestation","attest_author":"https://pith.science/pith/3JQDU2X5BXM2OCQ2IQ63GQDH3S/action/author_attestation","sign_citation":"https://pith.science/pith/3JQDU2X5BXM2OCQ2IQ63GQDH3S/action/citation_signature","submit_replication":"https://pith.science/pith/3JQDU2X5BXM2OCQ2IQ63GQDH3S/action/replication_record"}},"created_at":"2026-06-03T01:05:45.626427+00:00","updated_at":"2026-06-03T01:05:45.626427+00:00"}