{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:PEF2OU4TOXCBZH4PVJNZFZ3AAW","short_pith_number":"pith:PEF2OU4T","schema_version":"1.0","canonical_sha256":"790ba7539375c41c9f8faa5b92e76005a784d06da9a90c3ef002668a5aa1b086","source":{"kind":"arxiv","id":"2503.11926","version":1},"attestation_state":"computed","paper":{"title":"Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Aleksander Madry, Bowen Baker, David Farhi, Jakub Pachocki, Joost Huizinga, Leo Gao, Melody Y. Guan, Wojciech Zaremba, Zehao Dou","submitted_at":"2025-03-14T23:50:34Z","abstract_excerpt":"Mitigating reward hacking--where AI systems misbehave due to flaws or misspecifications in their learning objectives--remains a key challenge in constructing capable and aligned models. We show that we can monitor a frontier reasoning model, such as OpenAI o3-mini, for reward hacking in agentic coding environments by using another LLM that observes the model's chain-of-thought (CoT) reasoning. CoT monitoring can be far more effective than monitoring agent actions and outputs alone, and we further found that a LLM weaker than o3-mini, namely GPT-4o, can effectively monitor a stronger model. Bec"},"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":"2503.11926","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2025-03-14T23:50:34Z","cross_cats_sorted":[],"title_canon_sha256":"102c0c2ed72098ed7888ecdd28cfdac614ee8c0d65e205fcd64c54134857b2c7","abstract_canon_sha256":"2fbdfbb085b5adefc733d5cd9bc5ae7e80bb1de223cb9dc850eefb5d8ee9d4a4"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-21T07:19:08.436004Z","signature_b64":"M8Jz6aemLtsxc+4vvwzHxn8WUhtiNiglvZRXc1TZRdwoSbkoQyz0BdYEq6PE4MAx1UpVc+xJpBit8nZqmiPNDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"790ba7539375c41c9f8faa5b92e76005a784d06da9a90c3ef002668a5aa1b086","last_reissued_at":"2026-05-21T07:19:08.432765Z","signature_status":"signed_v1","first_computed_at":"2026-05-21T07:19:08.432765Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Aleksander Madry, Bowen Baker, David Farhi, Jakub Pachocki, Joost Huizinga, Leo Gao, Melody Y. Guan, Wojciech Zaremba, Zehao Dou","submitted_at":"2025-03-14T23:50:34Z","abstract_excerpt":"Mitigating reward hacking--where AI systems misbehave due to flaws or misspecifications in their learning objectives--remains a key challenge in constructing capable and aligned models. We show that we can monitor a frontier reasoning model, such as OpenAI o3-mini, for reward hacking in agentic coding environments by using another LLM that observes the model's chain-of-thought (CoT) reasoning. CoT monitoring can be far more effective than monitoring agent actions and outputs alone, and we further found that a LLM weaker than o3-mini, namely GPT-4o, can effectively monitor a stronger model. Bec"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.11926","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/2503.11926/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":"2503.11926","created_at":"2026-05-21T07:19:08.434242+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.11926v1","created_at":"2026-05-21T07:19:08.434242+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.11926","created_at":"2026-05-21T07:19:08.434242+00:00"},{"alias_kind":"pith_short_12","alias_value":"PEF2OU4TOXCB","created_at":"2026-05-21T07:19:08.434242+00:00"},{"alias_kind":"pith_short_16","alias_value":"PEF2OU4TOXCBZH4P","created_at":"2026-05-21T07:19:08.434242+00:00"},{"alias_kind":"pith_short_8","alias_value":"PEF2OU4T","created_at":"2026-05-21T07:19:08.434242+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":24,"internal_anchor_count":24,"sample":[{"citing_arxiv_id":"2511.21654","citing_title":"EvilGenie: A Reward Hacking Benchmark","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2605.20744","citing_title":"Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scale","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2605.21384","citing_title":"SpecBench: Measuring Reward Hacking in Long-Horizon Coding Agents","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2605.15377","citing_title":"Ensemble Monitoring for AI Control: Diverse Signals Outweigh More Compute","ref_index":41,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16198","citing_title":"Formal Methods Meet LLMs: Auditing, Monitoring, and Intervention for Compliance of Advanced AI Systems","ref_index":13,"is_internal_anchor":true},{"citing_arxiv_id":"2605.18549","citing_title":"Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamics","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2509.18127","citing_title":"Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2510.24941","citing_title":"Can Aha Moments Be Fake? Identifying True and Decorative Thinking Steps in Chain-of-Thought","ref_index":2,"is_internal_anchor":true},{"citing_arxiv_id":"2509.08827","citing_title":"A Survey of Reinforcement Learning for Large Reasoning Models","ref_index":30,"is_internal_anchor":true},{"citing_arxiv_id":"2511.16858","citing_title":"Investigating Test Overfitting on SWE-bench","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2601.03267","citing_title":"OpenAI GPT-5 System Card","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2512.21110","citing_title":"Beyond Context: Large Language Models' Failure to Grasp Users' Intent","ref_index":74,"is_internal_anchor":true},{"citing_arxiv_id":"2603.04474","citing_title":"From Spark to Fire: Modeling and Mitigating Error Cascades in LLM-Based Multi-Agent Collaboration","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2605.02398","citing_title":"The Compliance Trap: How Structural Constraints Degrade Frontier AI Metacognition Under Adversarial Pressure","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08715","citing_title":"AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems","ref_index":2,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12673","citing_title":"Do Androids Dream of Breaking the Game? Systematically Auditing AI Agent Benchmarks with BenchJack","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12746","citing_title":"CoT-Guard: Small Models for Strong Monitoring","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08715","citing_title":"AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems","ref_index":2,"is_internal_anchor":true},{"citing_arxiv_id":"2604.23488","citing_title":"Do Synthetic Trajectories Reflect Real Reward Hacking? A Systematic Study on Monitoring In-the-Wild Hacking in Code Generation","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2604.06427","citing_title":"The Depth Ceiling: On the Limits of Large Language Models in Discovering Latent Planning","ref_index":2,"is_internal_anchor":true},{"citing_arxiv_id":"2604.13602","citing_title":"Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges","ref_index":26,"is_internal_anchor":true},{"citing_arxiv_id":"2604.15149","citing_title":"LLMs Gaming Verifiers: RLVR can Lead to Reward Hacking","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2604.17761","citing_title":"Contrastive Attribution in the Wild: An Interpretability Analysis of LLM Failures on Realistic Benchmarks","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2605.02398","citing_title":"The Compliance Trap: How Structural Constraints Degrade Frontier AI Metacognition Under Adversarial Pressure","ref_index":1,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PEF2OU4TOXCBZH4PVJNZFZ3AAW","json":"https://pith.science/pith/PEF2OU4TOXCBZH4PVJNZFZ3AAW.json","graph_json":"https://pith.science/api/pith-number/PEF2OU4TOXCBZH4PVJNZFZ3AAW/graph.json","events_json":"https://pith.science/api/pith-number/PEF2OU4TOXCBZH4PVJNZFZ3AAW/events.json","paper":"https://pith.science/paper/PEF2OU4T"},"agent_actions":{"view_html":"https://pith.science/pith/PEF2OU4TOXCBZH4PVJNZFZ3AAW","download_json":"https://pith.science/pith/PEF2OU4TOXCBZH4PVJNZFZ3AAW.json","view_paper":"https://pith.science/paper/PEF2OU4T","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.11926&json=true","fetch_graph":"https://pith.science/api/pith-number/PEF2OU4TOXCBZH4PVJNZFZ3AAW/graph.json","fetch_events":"https://pith.science/api/pith-number/PEF2OU4TOXCBZH4PVJNZFZ3AAW/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PEF2OU4TOXCBZH4PVJNZFZ3AAW/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PEF2OU4TOXCBZH4PVJNZFZ3AAW/action/storage_attestation","attest_author":"https://pith.science/pith/PEF2OU4TOXCBZH4PVJNZFZ3AAW/action/author_attestation","sign_citation":"https://pith.science/pith/PEF2OU4TOXCBZH4PVJNZFZ3AAW/action/citation_signature","submit_replication":"https://pith.science/pith/PEF2OU4TOXCBZH4PVJNZFZ3AAW/action/replication_record"}},"created_at":"2026-05-21T07:19:08.434242+00:00","updated_at":"2026-05-21T07:19:08.434242+00:00"}