{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:QMYUDFWAX4FXXBHUYD2P2UWX3Z","short_pith_number":"pith:QMYUDFWA","schema_version":"1.0","canonical_sha256":"83314196c0bf0b7b84f4c0f4fd52d7de5c2fecc470f21faa21bd1108d00c7560","source":{"kind":"arxiv","id":"2506.22777","version":2},"attestation_state":"computed","paper":{"title":"Teaching Models to Verbalize Reward Hacking in Chain-of-Thought Reasoning","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Andy Arditi, Joe Benton, Julian Michael, Marvin Li, Miles Turpin","submitted_at":"2025-06-28T06:37:10Z","abstract_excerpt":"Language models trained with reinforcement learning (RL) can engage in reward hacking--the exploitation of unintended strategies for high reward--without revealing this behavior in their chain-of-thought reasoning. This makes the detection of reward hacking difficult, posing risks for high-stakes applications. We propose verbalization fine-tuning (VFT), a pre-RL fine-tuning intervention that trains models to explicitly acknowledge when they are influenced by prompt cues--hints which point to incorrect answers (e.g., \"a Stanford professor thinks the answer is A\"). To evaluate VFT, we subsequent"},"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":"2506.22777","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2025-06-28T06:37:10Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"4a99ca29dd93783ca280883a4e2b876f7f49751323e1ebdaf84df9f4d166d29f","abstract_canon_sha256":"d5c02fa9d9f6cf323e0ad8233d295fd38a3969fc9140e7f20bc933ea4299d83c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:36:33.746376Z","signature_b64":"y4TPQFHeHVzUZw37r5dgibNgzOBXus6IZxQCbTewQ2voQJcXDF24yvKWqF1+kBQmM32amBjHpb6gzjzHN2TLDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"83314196c0bf0b7b84f4c0f4fd52d7de5c2fecc470f21faa21bd1108d00c7560","last_reissued_at":"2026-07-05T11:36:33.745862Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:36:33.745862Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Teaching Models to Verbalize Reward Hacking in Chain-of-Thought Reasoning","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Andy Arditi, Joe Benton, Julian Michael, Marvin Li, Miles Turpin","submitted_at":"2025-06-28T06:37:10Z","abstract_excerpt":"Language models trained with reinforcement learning (RL) can engage in reward hacking--the exploitation of unintended strategies for high reward--without revealing this behavior in their chain-of-thought reasoning. This makes the detection of reward hacking difficult, posing risks for high-stakes applications. We propose verbalization fine-tuning (VFT), a pre-RL fine-tuning intervention that trains models to explicitly acknowledge when they are influenced by prompt cues--hints which point to incorrect answers (e.g., \"a Stanford professor thinks the answer is A\"). To evaluate VFT, we subsequent"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.22777","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/2506.22777/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":"2506.22777","created_at":"2026-07-05T11:36:33.745924+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.22777v2","created_at":"2026-07-05T11:36:33.745924+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.22777","created_at":"2026-07-05T11:36:33.745924+00:00"},{"alias_kind":"pith_short_12","alias_value":"QMYUDFWAX4FX","created_at":"2026-07-05T11:36:33.745924+00:00"},{"alias_kind":"pith_short_16","alias_value":"QMYUDFWAX4FXXBHU","created_at":"2026-07-05T11:36:33.745924+00:00"},{"alias_kind":"pith_short_8","alias_value":"QMYUDFWA","created_at":"2026-07-05T11:36:33.745924+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.24286","citing_title":"Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoning","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.25189","citing_title":"Directional Alignment Mitigates Reward Hacking in Reinforcement Learning for Language Models","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02964","citing_title":"Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2604.13602","citing_title":"Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges","ref_index":75,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/QMYUDFWAX4FXXBHUYD2P2UWX3Z","json":"https://pith.science/pith/QMYUDFWAX4FXXBHUYD2P2UWX3Z.json","graph_json":"https://pith.science/api/pith-number/QMYUDFWAX4FXXBHUYD2P2UWX3Z/graph.json","events_json":"https://pith.science/api/pith-number/QMYUDFWAX4FXXBHUYD2P2UWX3Z/events.json","paper":"https://pith.science/paper/QMYUDFWA"},"agent_actions":{"view_html":"https://pith.science/pith/QMYUDFWAX4FXXBHUYD2P2UWX3Z","download_json":"https://pith.science/pith/QMYUDFWAX4FXXBHUYD2P2UWX3Z.json","view_paper":"https://pith.science/paper/QMYUDFWA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.22777&json=true","fetch_graph":"https://pith.science/api/pith-number/QMYUDFWAX4FXXBHUYD2P2UWX3Z/graph.json","fetch_events":"https://pith.science/api/pith-number/QMYUDFWAX4FXXBHUYD2P2UWX3Z/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/QMYUDFWAX4FXXBHUYD2P2UWX3Z/action/timestamp_anchor","attest_storage":"https://pith.science/pith/QMYUDFWAX4FXXBHUYD2P2UWX3Z/action/storage_attestation","attest_author":"https://pith.science/pith/QMYUDFWAX4FXXBHUYD2P2UWX3Z/action/author_attestation","sign_citation":"https://pith.science/pith/QMYUDFWAX4FXXBHUYD2P2UWX3Z/action/citation_signature","submit_replication":"https://pith.science/pith/QMYUDFWAX4FXXBHUYD2P2UWX3Z/action/replication_record"}},"created_at":"2026-07-05T11:36:33.745924+00:00","updated_at":"2026-07-05T11:36:33.745924+00:00"}