{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:WST5PJHJHEHALSHLGFFZYOBQ5Y","short_pith_number":"pith:WST5PJHJ","schema_version":"1.0","canonical_sha256":"b4a7d7a4e9390e05c8eb314b9c3830ee2591233a47f160cc7dfc709e18dd4d3d","source":{"kind":"arxiv","id":"2403.07708","version":2},"attestation_state":"computed","paper":{"title":"Improving Reinforcement Learning from Human Feedback Using Contrastive Rewards","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Hongyi Guo, Rui Zheng, Wei Shen, Xiaoying Zhang, Yang Liu, Yuanshun Yao","submitted_at":"2024-03-12T14:51:57Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) is the mainstream paradigm used to align large language models (LLMs) with human preferences. Yet existing RLHF heavily relies on accurate and informative reward models, which are vulnerable and sensitive to noise from various sources, e.g. human labeling errors, making the pipeline fragile. In this work, we improve the effectiveness of the reward model by introducing a penalty term on the reward, named as \\textit{contrastive rewards}. %Contrastive rewards Our approach involves two steps: (1) an offline sampling step to obtain responses to prom"},"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":"2403.07708","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-03-12T14:51:57Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"67902a2a9e2ccfb4d0fae2e9a556d733fa88a945eab4c0e24a5ca9824537fe4b","abstract_canon_sha256":"3acc8d4032650468e95844502dd65bade1873d3c6cf8db09197821a0dc767396"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:55:46.723428Z","signature_b64":"pjKJFTZQqbXSKcL/4/VvzZYJhBgkdHcDslCm4bQWDu79wi09Bp3AFFm6fjyZZxJpvy80sLiZqw9rHA0DzEQzBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b4a7d7a4e9390e05c8eb314b9c3830ee2591233a47f160cc7dfc709e18dd4d3d","last_reissued_at":"2026-07-05T07:55:46.722891Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:55:46.722891Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Improving Reinforcement Learning from Human Feedback Using Contrastive Rewards","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Hongyi Guo, Rui Zheng, Wei Shen, Xiaoying Zhang, Yang Liu, Yuanshun Yao","submitted_at":"2024-03-12T14:51:57Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) is the mainstream paradigm used to align large language models (LLMs) with human preferences. Yet existing RLHF heavily relies on accurate and informative reward models, which are vulnerable and sensitive to noise from various sources, e.g. human labeling errors, making the pipeline fragile. In this work, we improve the effectiveness of the reward model by introducing a penalty term on the reward, named as \\textit{contrastive rewards}. %Contrastive rewards Our approach involves two steps: (1) an offline sampling step to obtain responses to prom"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2403.07708","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/2403.07708/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":"2403.07708","created_at":"2026-07-05T07:55:46.722953+00:00"},{"alias_kind":"arxiv_version","alias_value":"2403.07708v2","created_at":"2026-07-05T07:55:46.722953+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2403.07708","created_at":"2026-07-05T07:55:46.722953+00:00"},{"alias_kind":"pith_short_12","alias_value":"WST5PJHJHEHA","created_at":"2026-07-05T07:55:46.722953+00:00"},{"alias_kind":"pith_short_16","alias_value":"WST5PJHJHEHALSHL","created_at":"2026-07-05T07:55:46.722953+00:00"},{"alias_kind":"pith_short_8","alias_value":"WST5PJHJ","created_at":"2026-07-05T07:55:46.722953+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2504.02181","citing_title":"A Survey of Scaling in Large Language Model Reasoning","ref_index":180,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/WST5PJHJHEHALSHLGFFZYOBQ5Y","json":"https://pith.science/pith/WST5PJHJHEHALSHLGFFZYOBQ5Y.json","graph_json":"https://pith.science/api/pith-number/WST5PJHJHEHALSHLGFFZYOBQ5Y/graph.json","events_json":"https://pith.science/api/pith-number/WST5PJHJHEHALSHLGFFZYOBQ5Y/events.json","paper":"https://pith.science/paper/WST5PJHJ"},"agent_actions":{"view_html":"https://pith.science/pith/WST5PJHJHEHALSHLGFFZYOBQ5Y","download_json":"https://pith.science/pith/WST5PJHJHEHALSHLGFFZYOBQ5Y.json","view_paper":"https://pith.science/paper/WST5PJHJ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2403.07708&json=true","fetch_graph":"https://pith.science/api/pith-number/WST5PJHJHEHALSHLGFFZYOBQ5Y/graph.json","fetch_events":"https://pith.science/api/pith-number/WST5PJHJHEHALSHLGFFZYOBQ5Y/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/WST5PJHJHEHALSHLGFFZYOBQ5Y/action/timestamp_anchor","attest_storage":"https://pith.science/pith/WST5PJHJHEHALSHLGFFZYOBQ5Y/action/storage_attestation","attest_author":"https://pith.science/pith/WST5PJHJHEHALSHLGFFZYOBQ5Y/action/author_attestation","sign_citation":"https://pith.science/pith/WST5PJHJHEHALSHLGFFZYOBQ5Y/action/citation_signature","submit_replication":"https://pith.science/pith/WST5PJHJHEHALSHLGFFZYOBQ5Y/action/replication_record"}},"created_at":"2026-07-05T07:55:46.722953+00:00","updated_at":"2026-07-05T07:55:46.722953+00:00"}