{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:GRI5B4Q77J7RAYHLX6BFMH4LEU","short_pith_number":"pith:GRI5B4Q7","schema_version":"1.0","canonical_sha256":"3451d0f21ffa7f1060ebbf82561f8b250d934a1ee02c91541c122799a3d6eee9","source":{"kind":"arxiv","id":"2401.01879","version":3},"attestation_state":"computed","paper":{"title":"Theoretical guarantees on the best-of-n alignment policy","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.IT","math.IT"],"primary_cat":"cs.LG","authors_text":"Ahmad Beirami, Alekh Agarwal, Alexander D'Amour, Ananda Theertha Suresh, Chirag Nagpal, Jacob Eisenstein, Jonathan Berant","submitted_at":"2024-01-03T18:39:13Z","abstract_excerpt":"A simple and effective method for the inference-time alignment and scaling test-time compute of generative models is best-of-$n$ sampling, where $n$ samples are drawn from a reference policy, ranked based on a reward function, and the highest ranking one is selected. A commonly used analytical expression in the literature claims that the KL divergence between the best-of-$n$ policy and the reference policy is equal to $\\log (n) - (n-1)/n.$ We disprove the validity of this claim, and show that it is an upper bound on the actual KL divergence. We also explore the tightness of this upper bound in"},"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":"2401.01879","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-01-03T18:39:13Z","cross_cats_sorted":["cs.CL","cs.IT","math.IT"],"title_canon_sha256":"ae3d87fcff3fce86fe653e1946f577b15e47e1cef781f81523abfaaf3a514983","abstract_canon_sha256":"68a7625169e50d32484d64941048ac686030b2f5241ffc6812c7769798fb64f1"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:11:30.709554Z","signature_b64":"6uUg6PLC5hsOOcNh1wnfrvT9F+ywMbS72rym0PSUq9hjlzLLDTjsTfpo00OlXbcI+fyS8pU2HcflgH7WSGfgAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3451d0f21ffa7f1060ebbf82561f8b250d934a1ee02c91541c122799a3d6eee9","last_reissued_at":"2026-07-05T11:11:30.709019Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:11:30.709019Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Theoretical guarantees on the best-of-n alignment policy","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.IT","math.IT"],"primary_cat":"cs.LG","authors_text":"Ahmad Beirami, Alekh Agarwal, Alexander D'Amour, Ananda Theertha Suresh, Chirag Nagpal, Jacob Eisenstein, Jonathan Berant","submitted_at":"2024-01-03T18:39:13Z","abstract_excerpt":"A simple and effective method for the inference-time alignment and scaling test-time compute of generative models is best-of-$n$ sampling, where $n$ samples are drawn from a reference policy, ranked based on a reward function, and the highest ranking one is selected. A commonly used analytical expression in the literature claims that the KL divergence between the best-of-$n$ policy and the reference policy is equal to $\\log (n) - (n-1)/n.$ We disprove the validity of this claim, and show that it is an upper bound on the actual KL divergence. We also explore the tightness of this upper bound in"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2401.01879","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/2401.01879/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":"2401.01879","created_at":"2026-07-05T11:11:30.709084+00:00"},{"alias_kind":"arxiv_version","alias_value":"2401.01879v3","created_at":"2026-07-05T11:11:30.709084+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2401.01879","created_at":"2026-07-05T11:11:30.709084+00:00"},{"alias_kind":"pith_short_12","alias_value":"GRI5B4Q77J7R","created_at":"2026-07-05T11:11:30.709084+00:00"},{"alias_kind":"pith_short_16","alias_value":"GRI5B4Q77J7RAYHL","created_at":"2026-07-05T11:11:30.709084+00:00"},{"alias_kind":"pith_short_8","alias_value":"GRI5B4Q7","created_at":"2026-07-05T11:11:30.709084+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2505.15134","citing_title":"The Unreasonable Effectiveness of Entropy Minimization in LLM Reasoning","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14265","citing_title":"Reinforcement Learning via Value Gradient Flow","ref_index":6,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/GRI5B4Q77J7RAYHLX6BFMH4LEU","json":"https://pith.science/pith/GRI5B4Q77J7RAYHLX6BFMH4LEU.json","graph_json":"https://pith.science/api/pith-number/GRI5B4Q77J7RAYHLX6BFMH4LEU/graph.json","events_json":"https://pith.science/api/pith-number/GRI5B4Q77J7RAYHLX6BFMH4LEU/events.json","paper":"https://pith.science/paper/GRI5B4Q7"},"agent_actions":{"view_html":"https://pith.science/pith/GRI5B4Q77J7RAYHLX6BFMH4LEU","download_json":"https://pith.science/pith/GRI5B4Q77J7RAYHLX6BFMH4LEU.json","view_paper":"https://pith.science/paper/GRI5B4Q7","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2401.01879&json=true","fetch_graph":"https://pith.science/api/pith-number/GRI5B4Q77J7RAYHLX6BFMH4LEU/graph.json","fetch_events":"https://pith.science/api/pith-number/GRI5B4Q77J7RAYHLX6BFMH4LEU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/GRI5B4Q77J7RAYHLX6BFMH4LEU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/GRI5B4Q77J7RAYHLX6BFMH4LEU/action/storage_attestation","attest_author":"https://pith.science/pith/GRI5B4Q77J7RAYHLX6BFMH4LEU/action/author_attestation","sign_citation":"https://pith.science/pith/GRI5B4Q77J7RAYHLX6BFMH4LEU/action/citation_signature","submit_replication":"https://pith.science/pith/GRI5B4Q77J7RAYHLX6BFMH4LEU/action/replication_record"}},"created_at":"2026-07-05T11:11:30.709084+00:00","updated_at":"2026-07-05T11:11:30.709084+00:00"}