{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:VC6X3BEXBQLESEEOBZBKLAUL73","short_pith_number":"pith:VC6X3BEX","schema_version":"1.0","canonical_sha256":"a8bd7d84970c1649108e0e42a5828bfeca83eab7d67d4ae82d1e7f5ca775347d","source":{"kind":"arxiv","id":"2503.22230","version":3},"attestation_state":"computed","paper":{"title":"Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Chao Xin, Guanlin Liu, Lin Yan, Qingping Yang, Ruofei Zhu, Wei Shen, Yu Yue, Zheng Wu","submitted_at":"2025-03-28T08:26:41Z","abstract_excerpt":"Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning large language models with human preferences. While recent research has focused on algorithmic improvements, the importance of prompt-data construction has been overlooked. This paper addresses this gap by exploring data-driven bottlenecks in RLHF performance scaling, particularly reward hacking and decreasing response diversity. We introduce a hybrid reward system combining reasoning task verifiers (RTV) and a generative reward model (GenRM) to mitigate reward hacking. We also propose a novel prompt-selection method, Pr"},"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":"2503.22230","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-03-28T08:26:41Z","cross_cats_sorted":[],"title_canon_sha256":"9e57e31b4e817387e81a3855790a06a276dacc95bda4c9e0ee5f3c59775f0a50","abstract_canon_sha256":"be2fae74c0e3d4c1f3cb68f816e4cd582a46eef307a6f2bf32564318a119fcfa"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:43:27.026272Z","signature_b64":"DacWijrxlxsxxF3FL39AuqG/YQKhEuwhdtZ04kJ2lMlzqX6/hnvqr2iLQUFbjZQRz8rkevNZloTiJpYnAAFtBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a8bd7d84970c1649108e0e42a5828bfeca83eab7d67d4ae82d1e7f5ca775347d","last_reissued_at":"2026-07-05T10:43:27.025786Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:43:27.025786Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Chao Xin, Guanlin Liu, Lin Yan, Qingping Yang, Ruofei Zhu, Wei Shen, Yu Yue, Zheng Wu","submitted_at":"2025-03-28T08:26:41Z","abstract_excerpt":"Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning large language models with human preferences. While recent research has focused on algorithmic improvements, the importance of prompt-data construction has been overlooked. This paper addresses this gap by exploring data-driven bottlenecks in RLHF performance scaling, particularly reward hacking and decreasing response diversity. We introduce a hybrid reward system combining reasoning task verifiers (RTV) and a generative reward model (GenRM) to mitigate reward hacking. We also propose a novel prompt-selection method, Pr"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.22230","kind":"arxiv","version":3},"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/2503.22230/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":"2503.22230","created_at":"2026-07-05T10:43:27.025843+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.22230v3","created_at":"2026-07-05T10:43:27.025843+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.22230","created_at":"2026-07-05T10:43:27.025843+00:00"},{"alias_kind":"pith_short_12","alias_value":"VC6X3BEXBQLE","created_at":"2026-07-05T10:43:27.025843+00:00"},{"alias_kind":"pith_short_16","alias_value":"VC6X3BEXBQLESEEO","created_at":"2026-07-05T10:43:27.025843+00:00"},{"alias_kind":"pith_short_8","alias_value":"VC6X3BEX","created_at":"2026-07-05T10:43:27.025843+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2405.11143","citing_title":"OpenRLHF: An Easy-to-use, Scalable and High-performance RLHF Framework","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2504.05118","citing_title":"VAPO: Efficient and Reliable Reinforcement Learning for Advanced Reasoning Tasks","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12112","citing_title":"When Policy Entropy Constraint Fails: Preserving Diversity in Flow-based RLHF via Perceptual Entropy","ref_index":54,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08905","citing_title":"Forge: Quality-Aware Reinforcement Learning for NP-Hard Optimization in LLMs","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2505.07062","citing_title":"Seed1.5-VL Technical Report","ref_index":122,"is_internal_anchor":false},{"citing_arxiv_id":"2604.18381","citing_title":"Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes","ref_index":12,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/VC6X3BEXBQLESEEOBZBKLAUL73","json":"https://pith.science/pith/VC6X3BEXBQLESEEOBZBKLAUL73.json","graph_json":"https://pith.science/api/pith-number/VC6X3BEXBQLESEEOBZBKLAUL73/graph.json","events_json":"https://pith.science/api/pith-number/VC6X3BEXBQLESEEOBZBKLAUL73/events.json","paper":"https://pith.science/paper/VC6X3BEX"},"agent_actions":{"view_html":"https://pith.science/pith/VC6X3BEXBQLESEEOBZBKLAUL73","download_json":"https://pith.science/pith/VC6X3BEXBQLESEEOBZBKLAUL73.json","view_paper":"https://pith.science/paper/VC6X3BEX","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.22230&json=true","fetch_graph":"https://pith.science/api/pith-number/VC6X3BEXBQLESEEOBZBKLAUL73/graph.json","fetch_events":"https://pith.science/api/pith-number/VC6X3BEXBQLESEEOBZBKLAUL73/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VC6X3BEXBQLESEEOBZBKLAUL73/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VC6X3BEXBQLESEEOBZBKLAUL73/action/storage_attestation","attest_author":"https://pith.science/pith/VC6X3BEXBQLESEEOBZBKLAUL73/action/author_attestation","sign_citation":"https://pith.science/pith/VC6X3BEXBQLESEEOBZBKLAUL73/action/citation_signature","submit_replication":"https://pith.science/pith/VC6X3BEXBQLESEEOBZBKLAUL73/action/replication_record"}},"created_at":"2026-07-05T10:43:27.025843+00:00","updated_at":"2026-07-05T10:43:27.025843+00:00"}