{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:HL5YFUJSVDXGKH6F26QUUOY6MM","short_pith_number":"pith:HL5YFUJS","schema_version":"1.0","canonical_sha256":"3afb82d132a8ee651fc5d7a14a3b1e63114f83879d84de9eccd8cc460dae18d0","source":{"kind":"arxiv","id":"2504.06141","version":2},"attestation_state":"computed","paper":{"title":"Adversarial Training of Reward Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Adithya Renduchintala, Alexander Bukharin, Haifeng Qian, Oleksii Kuchaiev, Olivier Delalleau, Shengyang Sun, Soumye Singhal, Tuo Zhao, Zhilin Wang","submitted_at":"2025-04-08T15:38:25Z","abstract_excerpt":"Reward modeling has emerged as a promising approach for the scalable alignment of language models. However, contemporary reward models (RMs) often lack robustness, awarding high rewards to low-quality, out-of-distribution (OOD) samples. This can lead to reward hacking, where policies exploit unintended shortcuts to maximize rewards, undermining alignment. To address this challenge, we introduce Adv-RM, a novel adversarial training framework that automatically identifies adversarial examples -- responses that receive high rewards from the target RM but are OOD and of low quality. By leveraging "},"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":"2504.06141","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-04-08T15:38:25Z","cross_cats_sorted":[],"title_canon_sha256":"073f5a39139677eda19602a08a374341a7353f3683382dc0490c7a34784f78c8","abstract_canon_sha256":"6364a0b1bb21fce200f2844dc3e6cbe266252a11969831d695fe738dc87a1b9e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:47:57.957651Z","signature_b64":"u65yaBIpRlRxQLDP6VIymX9WI6rBRcIOKzsAn8krP/MnhuysNYtjyWK25xuG7ar5rpjgfonGpT5ZJOJQbD57BQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3afb82d132a8ee651fc5d7a14a3b1e63114f83879d84de9eccd8cc460dae18d0","last_reissued_at":"2026-07-05T10:47:57.957153Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:47:57.957153Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Adversarial Training of Reward Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Adithya Renduchintala, Alexander Bukharin, Haifeng Qian, Oleksii Kuchaiev, Olivier Delalleau, Shengyang Sun, Soumye Singhal, Tuo Zhao, Zhilin Wang","submitted_at":"2025-04-08T15:38:25Z","abstract_excerpt":"Reward modeling has emerged as a promising approach for the scalable alignment of language models. However, contemporary reward models (RMs) often lack robustness, awarding high rewards to low-quality, out-of-distribution (OOD) samples. This can lead to reward hacking, where policies exploit unintended shortcuts to maximize rewards, undermining alignment. To address this challenge, we introduce Adv-RM, a novel adversarial training framework that automatically identifies adversarial examples -- responses that receive high rewards from the target RM but are OOD and of low quality. By leveraging "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.06141","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/2504.06141/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":"2504.06141","created_at":"2026-07-05T10:47:57.957201+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.06141v2","created_at":"2026-07-05T10:47:57.957201+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.06141","created_at":"2026-07-05T10:47:57.957201+00:00"},{"alias_kind":"pith_short_12","alias_value":"HL5YFUJSVDXG","created_at":"2026-07-05T10:47:57.957201+00:00"},{"alias_kind":"pith_short_16","alias_value":"HL5YFUJSVDXGKH6F","created_at":"2026-07-05T10:47:57.957201+00:00"},{"alias_kind":"pith_short_8","alias_value":"HL5YFUJS","created_at":"2026-07-05T10:47:57.957201+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.09078","citing_title":"The Hidden Bias of Process Reward Models:PRISM for Rewarding the Right Reasoning","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2606.01249","citing_title":"Trust Region On-Policy Distillation","ref_index":151,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02686","citing_title":"Beyond Semantic Manipulation: Token-Space Attacks on Reward Models","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00155","citing_title":"Wasserstein Distributionally Robust Regret Optimization for Reinforcement Learning from Human Feedback","ref_index":4,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/HL5YFUJSVDXGKH6F26QUUOY6MM","json":"https://pith.science/pith/HL5YFUJSVDXGKH6F26QUUOY6MM.json","graph_json":"https://pith.science/api/pith-number/HL5YFUJSVDXGKH6F26QUUOY6MM/graph.json","events_json":"https://pith.science/api/pith-number/HL5YFUJSVDXGKH6F26QUUOY6MM/events.json","paper":"https://pith.science/paper/HL5YFUJS"},"agent_actions":{"view_html":"https://pith.science/pith/HL5YFUJSVDXGKH6F26QUUOY6MM","download_json":"https://pith.science/pith/HL5YFUJSVDXGKH6F26QUUOY6MM.json","view_paper":"https://pith.science/paper/HL5YFUJS","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.06141&json=true","fetch_graph":"https://pith.science/api/pith-number/HL5YFUJSVDXGKH6F26QUUOY6MM/graph.json","fetch_events":"https://pith.science/api/pith-number/HL5YFUJSVDXGKH6F26QUUOY6MM/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/HL5YFUJSVDXGKH6F26QUUOY6MM/action/timestamp_anchor","attest_storage":"https://pith.science/pith/HL5YFUJSVDXGKH6F26QUUOY6MM/action/storage_attestation","attest_author":"https://pith.science/pith/HL5YFUJSVDXGKH6F26QUUOY6MM/action/author_attestation","sign_citation":"https://pith.science/pith/HL5YFUJSVDXGKH6F26QUUOY6MM/action/citation_signature","submit_replication":"https://pith.science/pith/HL5YFUJSVDXGKH6F26QUUOY6MM/action/replication_record"}},"created_at":"2026-07-05T10:47:57.957201+00:00","updated_at":"2026-07-05T10:47:57.957201+00:00"}