{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:R6JSZODM3HOOYFMCLDDMKHBIZC","short_pith_number":"pith:R6JSZODM","schema_version":"1.0","canonical_sha256":"8f932cb86cd9dcec158258c6c51c28c8b5ed10c636d79f89eae48299346c15c6","source":{"kind":"arxiv","id":"2411.02306","version":3},"attestation_state":"computed","paper":{"title":"On Targeted Manipulation and Deception when Optimizing LLMs for User Feedback","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Adhyyan Narang, Anca Dragan, Brendan Murphy, Constantin Weisser, Marcus Williams, Micah Carroll","submitted_at":"2024-11-04T17:31:02Z","abstract_excerpt":"As LLMs become more widely deployed, there is increasing interest in directly optimizing for feedback from end users (e.g. thumbs up) in addition to feedback from paid annotators. However, training to maximize human feedback creates a perverse incentive structure for the AI to resort to manipulative or deceptive tactics to obtain positive feedback from users who are vulnerable to such strategies. We study this phenomenon by training LLMs with Reinforcement Learning with simulated user feedback in environments of practical LLM usage. In our settings, we find that: 1) Extreme forms of \"feedback "},"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":"2411.02306","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-11-04T17:31:02Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"c0fe04b6696a231d11584a81757d04450872321db79ed54fa1ba149e539675da","abstract_canon_sha256":"d3dd1fb07453a617d571cd2b719d45bcf2b1a40ca2005a6ed066d908a08abdb2"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:18:10.840857Z","signature_b64":"HGOE6vLUFFRh4ZXx1vl8DwXGl7msFtB6gvA8Vi+ZIkX9HzyEdgv57NAdBRq/OEOuKIh1+REs29EbgTDT8oq0AA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"8f932cb86cd9dcec158258c6c51c28c8b5ed10c636d79f89eae48299346c15c6","last_reissued_at":"2026-07-05T10:18:10.840381Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:18:10.840381Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"On Targeted Manipulation and Deception when Optimizing LLMs for User Feedback","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Adhyyan Narang, Anca Dragan, Brendan Murphy, Constantin Weisser, Marcus Williams, Micah Carroll","submitted_at":"2024-11-04T17:31:02Z","abstract_excerpt":"As LLMs become more widely deployed, there is increasing interest in directly optimizing for feedback from end users (e.g. thumbs up) in addition to feedback from paid annotators. However, training to maximize human feedback creates a perverse incentive structure for the AI to resort to manipulative or deceptive tactics to obtain positive feedback from users who are vulnerable to such strategies. We study this phenomenon by training LLMs with Reinforcement Learning with simulated user feedback in environments of practical LLM usage. In our settings, we find that: 1) Extreme forms of \"feedback "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2411.02306","kind":"arxiv","version":3},"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/2411.02306/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":"2411.02306","created_at":"2026-07-05T10:18:10.840439+00:00"},{"alias_kind":"arxiv_version","alias_value":"2411.02306v3","created_at":"2026-07-05T10:18:10.840439+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2411.02306","created_at":"2026-07-05T10:18:10.840439+00:00"},{"alias_kind":"pith_short_12","alias_value":"R6JSZODM3HOO","created_at":"2026-07-05T10:18:10.840439+00:00"},{"alias_kind":"pith_short_16","alias_value":"R6JSZODM3HOOYFMC","created_at":"2026-07-05T10:18:10.840439+00:00"},{"alias_kind":"pith_short_8","alias_value":"R6JSZODM","created_at":"2026-07-05T10:18:10.840439+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":10,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.23597","citing_title":"Against Proxy Optimization","ref_index":50,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05330","citing_title":"A Model of Multi-turn Human Persuadability Using Probabilistic Belief Tracing","ref_index":131,"is_internal_anchor":false},{"citing_arxiv_id":"2606.31916","citing_title":"Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action","ref_index":74,"is_internal_anchor":false},{"citing_arxiv_id":"2605.21602","citing_title":"Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs","ref_index":39,"is_internal_anchor":false},{"citing_arxiv_id":"2605.21602","citing_title":"Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs","ref_index":38,"is_internal_anchor":false},{"citing_arxiv_id":"2605.21602","citing_title":"Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs","ref_index":78,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09946","citing_title":"Structure from Strategic Interaction & Uncertainty: Risk Sensitive Games for Robust Preference Learning","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02585","citing_title":"Mitigating LLM biases toward spurious social contexts using direct preference optimization","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09808","citing_title":"Quantifying the Utility of User Simulators for Building Collaborative LLM Assistants","ref_index":93,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09946","citing_title":"Structure from Strategic Interaction & Uncertainty: Risk Sensitive Games for Robust Preference Learning","ref_index":2,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/R6JSZODM3HOOYFMCLDDMKHBIZC","json":"https://pith.science/pith/R6JSZODM3HOOYFMCLDDMKHBIZC.json","graph_json":"https://pith.science/api/pith-number/R6JSZODM3HOOYFMCLDDMKHBIZC/graph.json","events_json":"https://pith.science/api/pith-number/R6JSZODM3HOOYFMCLDDMKHBIZC/events.json","paper":"https://pith.science/paper/R6JSZODM"},"agent_actions":{"view_html":"https://pith.science/pith/R6JSZODM3HOOYFMCLDDMKHBIZC","download_json":"https://pith.science/pith/R6JSZODM3HOOYFMCLDDMKHBIZC.json","view_paper":"https://pith.science/paper/R6JSZODM","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2411.02306&json=true","fetch_graph":"https://pith.science/api/pith-number/R6JSZODM3HOOYFMCLDDMKHBIZC/graph.json","fetch_events":"https://pith.science/api/pith-number/R6JSZODM3HOOYFMCLDDMKHBIZC/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/R6JSZODM3HOOYFMCLDDMKHBIZC/action/timestamp_anchor","attest_storage":"https://pith.science/pith/R6JSZODM3HOOYFMCLDDMKHBIZC/action/storage_attestation","attest_author":"https://pith.science/pith/R6JSZODM3HOOYFMCLDDMKHBIZC/action/author_attestation","sign_citation":"https://pith.science/pith/R6JSZODM3HOOYFMCLDDMKHBIZC/action/citation_signature","submit_replication":"https://pith.science/pith/R6JSZODM3HOOYFMCLDDMKHBIZC/action/replication_record"}},"created_at":"2026-07-05T10:18:10.840439+00:00","updated_at":"2026-07-05T10:18:10.840439+00:00"}