{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:CQNGY75GNRUOLL4BH4QAMJWMB7","short_pith_number":"pith:CQNGY75G","schema_version":"1.0","canonical_sha256":"141a6c7fa66c68e5af813f200626cc0ff71c5d6ccf61e8dce7819361886bb514","source":{"kind":"arxiv","id":"2311.00168","version":2},"attestation_state":"computed","paper":{"title":"The Alignment Ceiling: Objective Mismatch in Reinforcement Learning from Human Feedback","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Nathan Lambert, Roberto Calandra","submitted_at":"2023-10-31T21:52:41Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) has emerged as a powerful technique to make large language models (LLMs) more capable in complex settings. RLHF proceeds as collecting human preference data, training a reward model on said data, and optimizing a base ML model with respect to said reward for extrinsic evaluation metrics (e.g. MMLU, GSM8k). RLHF relies on many assumptions about how the various pieces fit together, such as a reward model capturing human preferences and an RL optimizer extracting the right signal from a reward model. As the RLHF process involves many distinct desi"},"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":"2311.00168","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2023-10-31T21:52:41Z","cross_cats_sorted":[],"title_canon_sha256":"2caa4a523bcc1aec64da134b3652ec2359c7a321359c09057d364a80be5ad7d3","abstract_canon_sha256":"06565b32ae66bac1c13fb028fa2acf6f6697e6d320e70c0b0be0a4ee352eb5fb"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:40:24.356301Z","signature_b64":"3lv6MiYEfaGVCUnn6FjcmP7p3n3++FWZC3s4eZ5gMp5li4uQtv7SfCLQCEqp4d+3ZnwEWJOpPoLDBu4nAFlbBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"141a6c7fa66c68e5af813f200626cc0ff71c5d6ccf61e8dce7819361886bb514","last_reissued_at":"2026-07-05T07:40:24.355839Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:40:24.355839Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Alignment Ceiling: Objective Mismatch in Reinforcement Learning from Human Feedback","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Nathan Lambert, Roberto Calandra","submitted_at":"2023-10-31T21:52:41Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) has emerged as a powerful technique to make large language models (LLMs) more capable in complex settings. RLHF proceeds as collecting human preference data, training a reward model on said data, and optimizing a base ML model with respect to said reward for extrinsic evaluation metrics (e.g. MMLU, GSM8k). RLHF relies on many assumptions about how the various pieces fit together, such as a reward model capturing human preferences and an RL optimizer extracting the right signal from a reward model. As the RLHF process involves many distinct desi"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2311.00168","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/2311.00168/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":"2311.00168","created_at":"2026-07-05T07:40:24.355897+00:00"},{"alias_kind":"arxiv_version","alias_value":"2311.00168v2","created_at":"2026-07-05T07:40:24.355897+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2311.00168","created_at":"2026-07-05T07:40:24.355897+00:00"},{"alias_kind":"pith_short_12","alias_value":"CQNGY75GNRUO","created_at":"2026-07-05T07:40:24.355897+00:00"},{"alias_kind":"pith_short_16","alias_value":"CQNGY75GNRUOLL4B","created_at":"2026-07-05T07:40:24.355897+00:00"},{"alias_kind":"pith_short_8","alias_value":"CQNGY75G","created_at":"2026-07-05T07:40:24.355897+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.09711","citing_title":"Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization","ref_index":228,"is_internal_anchor":false},{"citing_arxiv_id":"2605.30323","citing_title":"In-Context Reward Adaptation for Robust Preference Modeling","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06674","citing_title":"What Do People Actually Want From AI? Mapping Preference Plurality","ref_index":55,"is_internal_anchor":false},{"citing_arxiv_id":"2402.05070","citing_title":"A Roadmap to Pluralistic Alignment","ref_index":195,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11974","citing_title":"Towards Order Fairness: Mitigating LLMs Order Sensitivity through Dual Group Advantage Optimization","ref_index":46,"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":60,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CQNGY75GNRUOLL4BH4QAMJWMB7","json":"https://pith.science/pith/CQNGY75GNRUOLL4BH4QAMJWMB7.json","graph_json":"https://pith.science/api/pith-number/CQNGY75GNRUOLL4BH4QAMJWMB7/graph.json","events_json":"https://pith.science/api/pith-number/CQNGY75GNRUOLL4BH4QAMJWMB7/events.json","paper":"https://pith.science/paper/CQNGY75G"},"agent_actions":{"view_html":"https://pith.science/pith/CQNGY75GNRUOLL4BH4QAMJWMB7","download_json":"https://pith.science/pith/CQNGY75GNRUOLL4BH4QAMJWMB7.json","view_paper":"https://pith.science/paper/CQNGY75G","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2311.00168&json=true","fetch_graph":"https://pith.science/api/pith-number/CQNGY75GNRUOLL4BH4QAMJWMB7/graph.json","fetch_events":"https://pith.science/api/pith-number/CQNGY75GNRUOLL4BH4QAMJWMB7/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CQNGY75GNRUOLL4BH4QAMJWMB7/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CQNGY75GNRUOLL4BH4QAMJWMB7/action/storage_attestation","attest_author":"https://pith.science/pith/CQNGY75GNRUOLL4BH4QAMJWMB7/action/author_attestation","sign_citation":"https://pith.science/pith/CQNGY75GNRUOLL4BH4QAMJWMB7/action/citation_signature","submit_replication":"https://pith.science/pith/CQNGY75GNRUOLL4BH4QAMJWMB7/action/replication_record"}},"created_at":"2026-07-05T07:40:24.355897+00:00","updated_at":"2026-07-05T07:40:24.355897+00:00"}