{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:F5H7G3CYIZ3ONQX4DHWZC4BGT4","short_pith_number":"pith:F5H7G3CY","schema_version":"1.0","canonical_sha256":"2f4ff36c584676e6c2fc19ed9170269f3aad63b3f156743cc8c5c3eecc9fde14","source":{"kind":"arxiv","id":"2505.03156","version":1},"attestation_state":"computed","paper":{"title":"Soft Best-of-n Sampling for Model Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","math.IT"],"primary_cat":"cs.IT","authors_text":"Alex Oesterling, Claudio Mayrink Verdun, Flavio P. Calmon, Himabindu Lakkaraju","submitted_at":"2025-05-06T04:03:11Z","abstract_excerpt":"Best-of-$n$ (BoN) sampling is a practical approach for aligning language model outputs with human preferences without expensive fine-tuning. BoN sampling is performed by generating $n$ responses to a prompt and then selecting the sample that maximizes a reward function. BoN yields high reward values in practice at a distortion cost, as measured by the KL-divergence between the sampled and original distribution. This distortion is coarsely controlled by varying the number of samples: larger $n$ yields a higher reward at a higher distortion cost. We introduce Soft Best-of-$n$ sampling, a general"},"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":"2505.03156","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.IT","submitted_at":"2025-05-06T04:03:11Z","cross_cats_sorted":["cs.AI","math.IT"],"title_canon_sha256":"27805f6ebf069ef404d0d78f5e0004a725126a625a49c1067d4359c3080a08ed","abstract_canon_sha256":"8406288a1eb6776e3bed064f1684d6b035a51e072ca7ff4d89786beec7ec2971"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:59:10.985686Z","signature_b64":"DQvWaPIYB5+TgcW0uDLYBrlppQUI33P47KauyoPwrvASNT0x16TnkYYckWSQRMp6vaH320nQTHJCU+yMG8DZCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2f4ff36c584676e6c2fc19ed9170269f3aad63b3f156743cc8c5c3eecc9fde14","last_reissued_at":"2026-07-05T10:59:10.985201Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:59:10.985201Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Soft Best-of-n Sampling for Model Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","math.IT"],"primary_cat":"cs.IT","authors_text":"Alex Oesterling, Claudio Mayrink Verdun, Flavio P. Calmon, Himabindu Lakkaraju","submitted_at":"2025-05-06T04:03:11Z","abstract_excerpt":"Best-of-$n$ (BoN) sampling is a practical approach for aligning language model outputs with human preferences without expensive fine-tuning. BoN sampling is performed by generating $n$ responses to a prompt and then selecting the sample that maximizes a reward function. BoN yields high reward values in practice at a distortion cost, as measured by the KL-divergence between the sampled and original distribution. This distortion is coarsely controlled by varying the number of samples: larger $n$ yields a higher reward at a higher distortion cost. We introduce Soft Best-of-$n$ sampling, a general"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.03156","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/2505.03156/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":"2505.03156","created_at":"2026-07-05T10:59:10.985259+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.03156v1","created_at":"2026-07-05T10:59:10.985259+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.03156","created_at":"2026-07-05T10:59:10.985259+00:00"},{"alias_kind":"pith_short_12","alias_value":"F5H7G3CYIZ3O","created_at":"2026-07-05T10:59:10.985259+00:00"},{"alias_kind":"pith_short_16","alias_value":"F5H7G3CYIZ3ONQX4","created_at":"2026-07-05T10:59:10.985259+00:00"},{"alias_kind":"pith_short_8","alias_value":"F5H7G3CY","created_at":"2026-07-05T10:59:10.985259+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2604.22212","citing_title":"Multimodal Diffusion to Mutually Enhance Polarized Light and Low Resolution EBSD Data","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01194","citing_title":"VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07105","citing_title":"Theoretical Limits of Language Model Alignment","ref_index":51,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/F5H7G3CYIZ3ONQX4DHWZC4BGT4","json":"https://pith.science/pith/F5H7G3CYIZ3ONQX4DHWZC4BGT4.json","graph_json":"https://pith.science/api/pith-number/F5H7G3CYIZ3ONQX4DHWZC4BGT4/graph.json","events_json":"https://pith.science/api/pith-number/F5H7G3CYIZ3ONQX4DHWZC4BGT4/events.json","paper":"https://pith.science/paper/F5H7G3CY"},"agent_actions":{"view_html":"https://pith.science/pith/F5H7G3CYIZ3ONQX4DHWZC4BGT4","download_json":"https://pith.science/pith/F5H7G3CYIZ3ONQX4DHWZC4BGT4.json","view_paper":"https://pith.science/paper/F5H7G3CY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.03156&json=true","fetch_graph":"https://pith.science/api/pith-number/F5H7G3CYIZ3ONQX4DHWZC4BGT4/graph.json","fetch_events":"https://pith.science/api/pith-number/F5H7G3CYIZ3ONQX4DHWZC4BGT4/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/F5H7G3CYIZ3ONQX4DHWZC4BGT4/action/timestamp_anchor","attest_storage":"https://pith.science/pith/F5H7G3CYIZ3ONQX4DHWZC4BGT4/action/storage_attestation","attest_author":"https://pith.science/pith/F5H7G3CYIZ3ONQX4DHWZC4BGT4/action/author_attestation","sign_citation":"https://pith.science/pith/F5H7G3CYIZ3ONQX4DHWZC4BGT4/action/citation_signature","submit_replication":"https://pith.science/pith/F5H7G3CYIZ3ONQX4DHWZC4BGT4/action/replication_record"}},"created_at":"2026-07-05T10:59:10.985259+00:00","updated_at":"2026-07-05T10:59:10.985259+00:00"}