{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:YXEY3DTOWRB4JSGMBTGFRIQATX","short_pith_number":"pith:YXEY3DTO","schema_version":"1.0","canonical_sha256":"c5c98d8e6eb443c4c8cc0ccc58a2009dd9213067635aeaaed30efcb7595efb66","source":{"kind":"arxiv","id":"2402.07319","version":1},"attestation_state":"computed","paper":{"title":"ODIN: Disentangled Reward Mitigates Hacking in RLHF","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Bryan Catanzaro, Chen Zhu, Davit Soselia, Heng Huang, Jiuhai Chen, Lichang Chen, Mohammad Shoeybi, Tianyi Zhou, Tom Goldstein","submitted_at":"2024-02-11T22:40:12Z","abstract_excerpt":"In this work, we study the issue of reward hacking on the response length, a challenge emerging in Reinforcement Learning from Human Feedback (RLHF) on LLMs. A well-formatted, verbose but less helpful response from the LLMs can often deceive LLMs or even human evaluators to achieve high scores. The same issue also holds for some reward models in RL. To address the challenges in both training and evaluation, we establish a more reliable evaluation protocol for comparing different training configurations, which inspects the trade-off between LLM evaluation score and response length obtained by v"},"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":"2402.07319","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-02-11T22:40:12Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"9bd25a5fdecb6033539a7c5e86ac23cd1a7e685cf72a8c7ea7549b3cc4f105ae","abstract_canon_sha256":"f8743a06e8a418f9701bf8db641cbbc67c6b9b5c5f733bf2c003b2ab676c2d2a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:44:04.509588Z","signature_b64":"nirPK5eThX54RfoJXZJZ/5NqYos9W1+kQoM+UARJL33dwk3cdlKJMRJnEXXPlRPIYwwBNNgu7uPT3GjF/4yBCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"c5c98d8e6eb443c4c8cc0ccc58a2009dd9213067635aeaaed30efcb7595efb66","last_reissued_at":"2026-07-05T07:44:04.509149Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:44:04.509149Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"ODIN: Disentangled Reward Mitigates Hacking in RLHF","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Bryan Catanzaro, Chen Zhu, Davit Soselia, Heng Huang, Jiuhai Chen, Lichang Chen, Mohammad Shoeybi, Tianyi Zhou, Tom Goldstein","submitted_at":"2024-02-11T22:40:12Z","abstract_excerpt":"In this work, we study the issue of reward hacking on the response length, a challenge emerging in Reinforcement Learning from Human Feedback (RLHF) on LLMs. A well-formatted, verbose but less helpful response from the LLMs can often deceive LLMs or even human evaluators to achieve high scores. The same issue also holds for some reward models in RL. To address the challenges in both training and evaluation, we establish a more reliable evaluation protocol for comparing different training configurations, which inspects the trade-off between LLM evaluation score and response length obtained by v"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2402.07319","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/2402.07319/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":"2402.07319","created_at":"2026-07-05T07:44:04.509211+00:00"},{"alias_kind":"arxiv_version","alias_value":"2402.07319v1","created_at":"2026-07-05T07:44:04.509211+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2402.07319","created_at":"2026-07-05T07:44:04.509211+00:00"},{"alias_kind":"pith_short_12","alias_value":"YXEY3DTOWRB4","created_at":"2026-07-05T07:44:04.509211+00:00"},{"alias_kind":"pith_short_16","alias_value":"YXEY3DTOWRB4JSGM","created_at":"2026-07-05T07:44:04.509211+00:00"},{"alias_kind":"pith_short_8","alias_value":"YXEY3DTO","created_at":"2026-07-05T07:44:04.509211+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":13,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2607.01830","citing_title":"Many Voices, One Reward: Multi-Role Rubric Generation for LLM Judging and Reward Modeling","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09711","citing_title":"Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization","ref_index":57,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18721","citing_title":"General Preference Reinforcement Learning","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17602","citing_title":"AutoRubric-T2I: Robust Rule-Based Reward Model for Text-to-Image Alignment","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2507.15698","citing_title":"CoLD: Counterfactually-Guided Length Debiasing for Process Reward Models in Mathematical Reasoning","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2601.21350","citing_title":"Factored Causal Representation Learning for Robust Reward Modeling in RLHF","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18721","citing_title":"General Preference Reinforcement Learning","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18721","citing_title":"General Preference Reinforcement Learning","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17602","citing_title":"AutoRubric-T2I: Robust Rule-Based Reward Model for Text-to-Image Alignment","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2506.12382","citing_title":"Exploring the Secondary Risks of Large Language Models","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2507.06419","citing_title":"Teach a Reward Model to Correct Itself: Reward Guided Adversarial Failure Discovery for Robust Reward Modeling","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2507.17746","citing_title":"Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05750","citing_title":"RVPO: Risk-Sensitive Alignment via Variance Regularization","ref_index":13,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/YXEY3DTOWRB4JSGMBTGFRIQATX","json":"https://pith.science/pith/YXEY3DTOWRB4JSGMBTGFRIQATX.json","graph_json":"https://pith.science/api/pith-number/YXEY3DTOWRB4JSGMBTGFRIQATX/graph.json","events_json":"https://pith.science/api/pith-number/YXEY3DTOWRB4JSGMBTGFRIQATX/events.json","paper":"https://pith.science/paper/YXEY3DTO"},"agent_actions":{"view_html":"https://pith.science/pith/YXEY3DTOWRB4JSGMBTGFRIQATX","download_json":"https://pith.science/pith/YXEY3DTOWRB4JSGMBTGFRIQATX.json","view_paper":"https://pith.science/paper/YXEY3DTO","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2402.07319&json=true","fetch_graph":"https://pith.science/api/pith-number/YXEY3DTOWRB4JSGMBTGFRIQATX/graph.json","fetch_events":"https://pith.science/api/pith-number/YXEY3DTOWRB4JSGMBTGFRIQATX/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/YXEY3DTOWRB4JSGMBTGFRIQATX/action/timestamp_anchor","attest_storage":"https://pith.science/pith/YXEY3DTOWRB4JSGMBTGFRIQATX/action/storage_attestation","attest_author":"https://pith.science/pith/YXEY3DTOWRB4JSGMBTGFRIQATX/action/author_attestation","sign_citation":"https://pith.science/pith/YXEY3DTOWRB4JSGMBTGFRIQATX/action/citation_signature","submit_replication":"https://pith.science/pith/YXEY3DTOWRB4JSGMBTGFRIQATX/action/replication_record"}},"created_at":"2026-07-05T07:44:04.509211+00:00","updated_at":"2026-07-05T07:44:04.509211+00:00"}