{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:E3VDQ5S76MEE4A3DWQYTNEJ6SH","short_pith_number":"pith:E3VDQ5S7","schema_version":"1.0","canonical_sha256":"26ea38765ff3084e0363b43136913e91e7e9e6866df4e211a0f9421dac716ff2","source":{"kind":"arxiv","id":"2506.08266","version":1},"attestation_state":"computed","paper":{"title":"Reinforcement Learning from Human Feedback with High-Confidence Safety Constraints","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","stat.AP"],"primary_cat":"cs.LG","authors_text":"Austin Hoag, Blossom Metevier, Philip S. Thomas, Scott Niekum, Will Schwarzer, Yaswanth Chittepu","submitted_at":"2025-06-09T22:03:56Z","abstract_excerpt":"Existing approaches to language model alignment often treat safety as a tradeoff against helpfulness, which can lead to unacceptable responses in sensitive domains. To ensure reliable performance in such settings, we propose High-Confidence Safe Reinforcement Learning from Human Feedback (HC-RLHF), a method that provides high-confidence safety guarantees while maximizing helpfulness. Similar to previous methods, HC-RLHF explicitly decouples human preferences into helpfulness and harmlessness (safety), which are learned by training a reward model and a cost model, respectively. It then employs "},"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":"2506.08266","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-06-09T22:03:56Z","cross_cats_sorted":["cs.AI","cs.CL","stat.AP"],"title_canon_sha256":"06693bab1075ada37e93ef2686655ea9fd294f6722299c9f2edd28c3cc94a690","abstract_canon_sha256":"c4fdfde24d51eebe4b1ede64a8727e934e1f527bb3c2335b48637dc3e486a10a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:18:47.185421Z","signature_b64":"dU1RqoPudM3KHKeAGc2ZI2op2gMO0+iX/iF2Mmn21s6/dgv/hdTR/tZUgXb/aM6KQHwCWCRAICFU6Jk4bZ0oDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"26ea38765ff3084e0363b43136913e91e7e9e6866df4e211a0f9421dac716ff2","last_reissued_at":"2026-07-05T11:18:47.184938Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:18:47.184938Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Reinforcement Learning from Human Feedback with High-Confidence Safety Constraints","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","stat.AP"],"primary_cat":"cs.LG","authors_text":"Austin Hoag, Blossom Metevier, Philip S. Thomas, Scott Niekum, Will Schwarzer, Yaswanth Chittepu","submitted_at":"2025-06-09T22:03:56Z","abstract_excerpt":"Existing approaches to language model alignment often treat safety as a tradeoff against helpfulness, which can lead to unacceptable responses in sensitive domains. To ensure reliable performance in such settings, we propose High-Confidence Safe Reinforcement Learning from Human Feedback (HC-RLHF), a method that provides high-confidence safety guarantees while maximizing helpfulness. Similar to previous methods, HC-RLHF explicitly decouples human preferences into helpfulness and harmlessness (safety), which are learned by training a reward model and a cost model, respectively. It then employs "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.08266","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/2506.08266/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":"2506.08266","created_at":"2026-07-05T11:18:47.185004+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.08266v1","created_at":"2026-07-05T11:18:47.185004+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.08266","created_at":"2026-07-05T11:18:47.185004+00:00"},{"alias_kind":"pith_short_12","alias_value":"E3VDQ5S76MEE","created_at":"2026-07-05T11:18:47.185004+00:00"},{"alias_kind":"pith_short_16","alias_value":"E3VDQ5S76MEE4A3D","created_at":"2026-07-05T11:18:47.185004+00:00"},{"alias_kind":"pith_short_8","alias_value":"E3VDQ5S7","created_at":"2026-07-05T11:18:47.185004+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.21822","citing_title":"Implicit Safety Alignment from Crowd Preferences","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02507","citing_title":"Reinforcement Learning from Human Feedback: A Statistical Perspective","ref_index":14,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/E3VDQ5S76MEE4A3DWQYTNEJ6SH","json":"https://pith.science/pith/E3VDQ5S76MEE4A3DWQYTNEJ6SH.json","graph_json":"https://pith.science/api/pith-number/E3VDQ5S76MEE4A3DWQYTNEJ6SH/graph.json","events_json":"https://pith.science/api/pith-number/E3VDQ5S76MEE4A3DWQYTNEJ6SH/events.json","paper":"https://pith.science/paper/E3VDQ5S7"},"agent_actions":{"view_html":"https://pith.science/pith/E3VDQ5S76MEE4A3DWQYTNEJ6SH","download_json":"https://pith.science/pith/E3VDQ5S76MEE4A3DWQYTNEJ6SH.json","view_paper":"https://pith.science/paper/E3VDQ5S7","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.08266&json=true","fetch_graph":"https://pith.science/api/pith-number/E3VDQ5S76MEE4A3DWQYTNEJ6SH/graph.json","fetch_events":"https://pith.science/api/pith-number/E3VDQ5S76MEE4A3DWQYTNEJ6SH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/E3VDQ5S76MEE4A3DWQYTNEJ6SH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/E3VDQ5S76MEE4A3DWQYTNEJ6SH/action/storage_attestation","attest_author":"https://pith.science/pith/E3VDQ5S76MEE4A3DWQYTNEJ6SH/action/author_attestation","sign_citation":"https://pith.science/pith/E3VDQ5S76MEE4A3DWQYTNEJ6SH/action/citation_signature","submit_replication":"https://pith.science/pith/E3VDQ5S76MEE4A3DWQYTNEJ6SH/action/replication_record"}},"created_at":"2026-07-05T11:18:47.185004+00:00","updated_at":"2026-07-05T11:18:47.185004+00:00"}