{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:IHWICRQNIDMAKYSYRTSSWSTWRY","short_pith_number":"pith:IHWICRQN","schema_version":"1.0","canonical_sha256":"41ec81460d40d80562588ce52b4a768e12487461a528522a7e07c7de9071a6f4","source":{"kind":"arxiv","id":"2309.09117","version":2},"attestation_state":"computed","paper":{"title":"Contrastive Decoding Improves Reasoning in Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Mike Lewis, Sean O'Brien","submitted_at":"2023-09-17T00:29:32Z","abstract_excerpt":"We demonstrate that Contrastive Decoding -- a simple, computationally light, and training-free text generation method proposed by Li et al 2022 -- achieves large out-of-the-box improvements over greedy decoding on a variety of reasoning tasks. Originally shown to improve the perceived quality of long-form text generation, Contrastive Decoding searches for strings that maximize a weighted difference in likelihood between strong and weak models. We show that Contrastive Decoding leads LLaMA-65B to outperform LLaMA 2, GPT-3.5 and PaLM 2-L on the HellaSwag commonsense reasoning benchmark, and to o"},"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":"2309.09117","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2023-09-17T00:29:32Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"04019e91050893bfa4cac58154dba7e32e243a609f1fb321e6c769bd2dc6da60","abstract_canon_sha256":"7e486fb11762774946db5d13a4c7a97db5403594e4fc3a447ea4808d3881cf1f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T06:55:40.575476Z","signature_b64":"TSiH834SbXljk8k/pfTIG4Q2XYwgPjZF0NefOVy/xlQn6k27Gx1xloXqglKp7n46R2ERV0ePB8ySw9ETGDsLBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"41ec81460d40d80562588ce52b4a768e12487461a528522a7e07c7de9071a6f4","last_reissued_at":"2026-07-05T06:55:40.574981Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T06:55:40.574981Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Contrastive Decoding Improves Reasoning in Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Mike Lewis, Sean O'Brien","submitted_at":"2023-09-17T00:29:32Z","abstract_excerpt":"We demonstrate that Contrastive Decoding -- a simple, computationally light, and training-free text generation method proposed by Li et al 2022 -- achieves large out-of-the-box improvements over greedy decoding on a variety of reasoning tasks. Originally shown to improve the perceived quality of long-form text generation, Contrastive Decoding searches for strings that maximize a weighted difference in likelihood between strong and weak models. We show that Contrastive Decoding leads LLaMA-65B to outperform LLaMA 2, GPT-3.5 and PaLM 2-L on the HellaSwag commonsense reasoning benchmark, and to o"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2309.09117","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/2309.09117/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":"2309.09117","created_at":"2026-07-05T06:55:40.575038+00:00"},{"alias_kind":"arxiv_version","alias_value":"2309.09117v2","created_at":"2026-07-05T06:55:40.575038+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2309.09117","created_at":"2026-07-05T06:55:40.575038+00:00"},{"alias_kind":"pith_short_12","alias_value":"IHWICRQNIDMA","created_at":"2026-07-05T06:55:40.575038+00:00"},{"alias_kind":"pith_short_16","alias_value":"IHWICRQNIDMAKYSY","created_at":"2026-07-05T06:55:40.575038+00:00"},{"alias_kind":"pith_short_8","alias_value":"IHWICRQN","created_at":"2026-07-05T06:55:40.575038+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":12,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.26080","citing_title":"Neglected Free Lunch from Post-training: Progress Advantage for LLM Agents","ref_index":121,"is_internal_anchor":false},{"citing_arxiv_id":"2606.01923","citing_title":"Resonant Context Anchoring: Decoupling Attention Routing and Signal Gain at Inference Time","ref_index":71,"is_internal_anchor":false},{"citing_arxiv_id":"2606.01682","citing_title":"Off-the-Shelf LLMs as Process Scorers: Training-Free Alternative to PRMs for Mathematical Reasoning","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2605.25902","citing_title":"Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding Diffing","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.29198","citing_title":"Guidance Contrastive Token Credit Assignment for Discrete Policy Optimization","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2605.31429","citing_title":"YARD: Y-Architecture Register Decoding for Efficient Hallucination Mitigation in Large Vision-Language Models","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2510.07962","citing_title":"LightReasoner: Can Small Language Models Teach Large Language Models Reasoning?","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2309.03883","citing_title":"DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models","ref_index":91,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12185","citing_title":"Mitigating Context-Memory Conflicts in LLMs through Dynamic Cognitive Reconciliation Decoding","ref_index":38,"is_internal_anchor":false},{"citing_arxiv_id":"2311.05232","citing_title":"A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions","ref_index":240,"is_internal_anchor":false},{"citing_arxiv_id":"2604.18106","citing_title":"Efficient Low-Resource Language Adaptation via Multi-Source Dynamic Logit Fusion","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19124","citing_title":"Detoxification for LLM: From Dataset Itself","ref_index":4,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/IHWICRQNIDMAKYSYRTSSWSTWRY","json":"https://pith.science/pith/IHWICRQNIDMAKYSYRTSSWSTWRY.json","graph_json":"https://pith.science/api/pith-number/IHWICRQNIDMAKYSYRTSSWSTWRY/graph.json","events_json":"https://pith.science/api/pith-number/IHWICRQNIDMAKYSYRTSSWSTWRY/events.json","paper":"https://pith.science/paper/IHWICRQN"},"agent_actions":{"view_html":"https://pith.science/pith/IHWICRQNIDMAKYSYRTSSWSTWRY","download_json":"https://pith.science/pith/IHWICRQNIDMAKYSYRTSSWSTWRY.json","view_paper":"https://pith.science/paper/IHWICRQN","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2309.09117&json=true","fetch_graph":"https://pith.science/api/pith-number/IHWICRQNIDMAKYSYRTSSWSTWRY/graph.json","fetch_events":"https://pith.science/api/pith-number/IHWICRQNIDMAKYSYRTSSWSTWRY/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/IHWICRQNIDMAKYSYRTSSWSTWRY/action/timestamp_anchor","attest_storage":"https://pith.science/pith/IHWICRQNIDMAKYSYRTSSWSTWRY/action/storage_attestation","attest_author":"https://pith.science/pith/IHWICRQNIDMAKYSYRTSSWSTWRY/action/author_attestation","sign_citation":"https://pith.science/pith/IHWICRQNIDMAKYSYRTSSWSTWRY/action/citation_signature","submit_replication":"https://pith.science/pith/IHWICRQNIDMAKYSYRTSSWSTWRY/action/replication_record"}},"created_at":"2026-07-05T06:55:40.575038+00:00","updated_at":"2026-07-05T06:55:40.575038+00:00"}