{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:BVKBZQUYSMOBMQBKNP7HC2KJRA","short_pith_number":"pith:BVKBZQUY","schema_version":"1.0","canonical_sha256":"0d541cc298931c16402a6bfe716949880f20c0f94d45e9ed8c96aeb23eb76ef6","source":{"kind":"arxiv","id":"2508.05928","version":1},"attestation_state":"computed","paper":{"title":"Mitigating Think-Answer Mismatch in LLM Reasoning Through Noise-Aware Advantage Reweighting","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Danhao Zhu, Peijun Shen, Si Shen, Wenhua Zhao","submitted_at":"2025-08-08T01:24:06Z","abstract_excerpt":"Group-Relative Policy Optimization (GRPO) is a key technique for training large reasoning models, yet it suffers from a critical vulnerability: the \\emph{Think-Answer Mismatch}, where noisy reward signals corrupt the learning process. This problem is most severe in unbalanced response groups, paradoxically degrading the signal precisely when it should be most informative. To address this challenge, we propose Stable Group-Relative Policy Optimization (S-GRPO), a principled enhancement that derives optimal, noise-aware advantage weights to stabilize training. Our comprehensive experiments on ma"},"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":"2508.05928","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-08-08T01:24:06Z","cross_cats_sorted":[],"title_canon_sha256":"09b3b0a6f5a9dd6a96ee79aa1f398205d4de7f7d900cdacccda10345afa46c25","abstract_canon_sha256":"df3446ed9f9f65f03fb1a822cc7868d446607b6f44ce620292f5f212f6ff69af"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:50:45.680948Z","signature_b64":"NgauonJXd8bKu7N0uSc1Wx1OSIEyW0i+L4xfoo59NXdSsbEzp0xChDpngNV9gOjoDnQf0QSNjPpuikGwhE7GAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0d541cc298931c16402a6bfe716949880f20c0f94d45e9ed8c96aeb23eb76ef6","last_reissued_at":"2026-07-05T11:50:45.680453Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:50:45.680453Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Mitigating Think-Answer Mismatch in LLM Reasoning Through Noise-Aware Advantage Reweighting","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Danhao Zhu, Peijun Shen, Si Shen, Wenhua Zhao","submitted_at":"2025-08-08T01:24:06Z","abstract_excerpt":"Group-Relative Policy Optimization (GRPO) is a key technique for training large reasoning models, yet it suffers from a critical vulnerability: the \\emph{Think-Answer Mismatch}, where noisy reward signals corrupt the learning process. This problem is most severe in unbalanced response groups, paradoxically degrading the signal precisely when it should be most informative. To address this challenge, we propose Stable Group-Relative Policy Optimization (S-GRPO), a principled enhancement that derives optimal, noise-aware advantage weights to stabilize training. Our comprehensive experiments on ma"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2508.05928","kind":"arxiv","version":1},"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/2508.05928/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":"2508.05928","created_at":"2026-07-05T11:50:45.680512+00:00"},{"alias_kind":"arxiv_version","alias_value":"2508.05928v1","created_at":"2026-07-05T11:50:45.680512+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2508.05928","created_at":"2026-07-05T11:50:45.680512+00:00"},{"alias_kind":"pith_short_12","alias_value":"BVKBZQUYSMOB","created_at":"2026-07-05T11:50:45.680512+00:00"},{"alias_kind":"pith_short_16","alias_value":"BVKBZQUYSMOBMQBK","created_at":"2026-07-05T11:50:45.680512+00:00"},{"alias_kind":"pith_short_8","alias_value":"BVKBZQUY","created_at":"2026-07-05T11:50:45.680512+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2511.11113","citing_title":"VIDEOP2R: Video Understanding from Perception to Reasoning","ref_index":44,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10863","citing_title":"DGPO: Beyond Pairwise Preferences with Directional Consistent Groupwise Optimization","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06755","citing_title":"Gradient Extrapolation-Based Policy Optimization","ref_index":29,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BVKBZQUYSMOBMQBKNP7HC2KJRA","json":"https://pith.science/pith/BVKBZQUYSMOBMQBKNP7HC2KJRA.json","graph_json":"https://pith.science/api/pith-number/BVKBZQUYSMOBMQBKNP7HC2KJRA/graph.json","events_json":"https://pith.science/api/pith-number/BVKBZQUYSMOBMQBKNP7HC2KJRA/events.json","paper":"https://pith.science/paper/BVKBZQUY"},"agent_actions":{"view_html":"https://pith.science/pith/BVKBZQUYSMOBMQBKNP7HC2KJRA","download_json":"https://pith.science/pith/BVKBZQUYSMOBMQBKNP7HC2KJRA.json","view_paper":"https://pith.science/paper/BVKBZQUY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2508.05928&json=true","fetch_graph":"https://pith.science/api/pith-number/BVKBZQUYSMOBMQBKNP7HC2KJRA/graph.json","fetch_events":"https://pith.science/api/pith-number/BVKBZQUYSMOBMQBKNP7HC2KJRA/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BVKBZQUYSMOBMQBKNP7HC2KJRA/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BVKBZQUYSMOBMQBKNP7HC2KJRA/action/storage_attestation","attest_author":"https://pith.science/pith/BVKBZQUYSMOBMQBKNP7HC2KJRA/action/author_attestation","sign_citation":"https://pith.science/pith/BVKBZQUYSMOBMQBKNP7HC2KJRA/action/citation_signature","submit_replication":"https://pith.science/pith/BVKBZQUYSMOBMQBKNP7HC2KJRA/action/replication_record"}},"created_at":"2026-07-05T11:50:45.680512+00:00","updated_at":"2026-07-05T11:50:45.680512+00:00"}