{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2023:LLVGQOES5C76SM7ACCGGBGCDKR","short_pith_number":"pith:LLVGQOES","canonical_record":{"source":{"id":"2309.03883","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2023-09-07T17:45:31Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"77b009683fd9984586fc03270635ba494e0f7dd2918c8434f39e8db6744e6858","abstract_canon_sha256":"501a75d39efa00a3933da886f01a9789c24c31af27d06ce55db1d4a2f828877e"},"schema_version":"1.0"},"canonical_sha256":"5aea683892e8bfe933e0108c609843544083e05d09431fbedf3c4cd1de5129a6","source":{"kind":"arxiv","id":"2309.03883","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2309.03883","created_at":"2026-05-17T23:38:47Z"},{"alias_kind":"arxiv_version","alias_value":"2309.03883v2","created_at":"2026-05-17T23:38:47Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2309.03883","created_at":"2026-05-17T23:38:47Z"},{"alias_kind":"pith_short_12","alias_value":"LLVGQOES5C76","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"LLVGQOES5C76SM7A","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"LLVGQOES","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2023:LLVGQOES5C76SM7ACCGGBGCDKR","target":"record","payload":{"canonical_record":{"source":{"id":"2309.03883","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2023-09-07T17:45:31Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"77b009683fd9984586fc03270635ba494e0f7dd2918c8434f39e8db6744e6858","abstract_canon_sha256":"501a75d39efa00a3933da886f01a9789c24c31af27d06ce55db1d4a2f828877e"},"schema_version":"1.0"},"canonical_sha256":"5aea683892e8bfe933e0108c609843544083e05d09431fbedf3c4cd1de5129a6","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:47.753260Z","signature_b64":"YKeeTCnp/2XSvGclevvbdXdNCEfC5+f0DUu1dyQKdZlqO1W23sv9v44FtZAiRQMERHeuXxDaW/aerc0/FuRVAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5aea683892e8bfe933e0108c609843544083e05d09431fbedf3c4cd1de5129a6","last_reissued_at":"2026-05-17T23:38:47.752809Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:47.752809Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2309.03883","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:47Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"TWOPCRhVc6yR0aprBbDMncUD0qVUFvd3J5WC2xCnz/OuLDfUdfsVKjEFCPjrYb8hPlRsiP6uhqtZOMtBKqIYBQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-30T15:40:21.235097Z"},"content_sha256":"264088566689ff28003910479cd7ff960b38b61bdeaad7576b26ed6daa11c96a","schema_version":"1.0","event_id":"sha256:264088566689ff28003910479cd7ff960b38b61bdeaad7576b26ed6daa11c96a"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2023:LLVGQOES5C76SM7ACCGGBGCDKR","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Contrasting logits from later versus earlier transformer layers reduces hallucinations in large language models.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Hongyin Luo, James Glass, Pengcheng He, Yoon Kim, Yujia Xie, Yung-Sung Chuang","submitted_at":"2023-09-07T17:45:31Z","abstract_excerpt":"Despite their impressive capabilities, large language models (LLMs) are prone to hallucinations, i.e., generating content that deviates from facts seen during pretraining. We propose a simple decoding strategy for reducing hallucinations with pretrained LLMs that does not require conditioning on retrieved external knowledge nor additional fine-tuning. Our approach obtains the next-token distribution by contrasting the differences in logits obtained from projecting the later layers versus earlier layers to the vocabulary space, exploiting the fact that factual knowledge in an LLMs has generally"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"DoLa consistently improves the truthfulness across multiple choices tasks and open-ended generation tasks, for example improving the performance of LLaMA family models on TruthfulQA by 12-17% absolute points.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The central assumption is that factual knowledge is localized to particular transformer layers and that subtracting early-layer logits from late-layer logits reliably surfaces that knowledge rather than introducing new errors.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DoLa reduces hallucinations in LLMs by contrasting logits from later versus earlier layers during decoding, improving truthfulness on TruthfulQA by 12-17 absolute points without fine-tuning or retrieval.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Contrasting logits from later versus earlier transformer layers reduces hallucinations in large language models.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d038f5d4d5b24ebeec67bd882d5e9ab4d37983fd64f32eae1c52530de9267a21"},"source":{"id":"2309.03883","kind":"arxiv","version":2},"verdict":{"id":"1da1cc43-1f1d-4bd4-a90e-611d5fe5f61b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T13:17:06.413596Z","strongest_claim":"DoLa consistently improves the truthfulness across multiple choices tasks and open-ended generation tasks, for example improving the performance of LLaMA family models on TruthfulQA by 12-17% absolute points.","one_line_summary":"DoLa reduces hallucinations in LLMs by contrasting logits from later versus earlier layers during decoding, improving truthfulness on TruthfulQA by 12-17 absolute points without fine-tuning or retrieval.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The central assumption is that factual knowledge is localized to particular transformer layers and that subtracting early-layer logits from late-layer logits reliably surfaces that knowledge rather than introducing new errors.","pith_extraction_headline":"Contrasting logits from later versus earlier transformer layers reduces hallucinations in large language models."},"references":{"count":77,"sample":[{"doi":"","year":2022,"title":"Improving language models by retrieving from trillions of tokens","work_id":"fef1f97e-3516-4e29-ac79-692015f03cc7","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1901,"title":"Language models are few-shot learners","work_id":"33f1f5a3-ad49-41fa-b313-857dabbce8e9","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Gonzalez, Ion Stoica, and Eric P","work_id":"0aa2a3b1-ff5f-48a1-8c2e-3e0efa840d1e","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Knowledge neurons in pretrained transformers","work_id":"f36915df-31b1-4f5f-8215-1b98fae10b02","ref_index":7,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"Depth-adaptive transformer","work_id":"a9fd354d-c8f4-4333-97fa-86e7481ac861","ref_index":9,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":77,"snapshot_sha256":"745ce7d68ed976e9178d55c4f94944e3c609ac69b899dd6aeecb2741c96ef91a","internal_anchors":11},"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"},"verdict_id":"1da1cc43-1f1d-4bd4-a90e-611d5fe5f61b"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:47Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"rF8SVZvZ1EupdCe+Z3oSpxGJc8F2hp/lnfYjGWL+1ER042bLVLRJUladvg9Ix4zLIR0apqyt66loRyiuqIanAQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-30T15:40:21.236046Z"},"content_sha256":"f3374e6a26a9d27a5cfa588d45337c014e42b411f1601bd0e9a6cb5518c6eef9","schema_version":"1.0","event_id":"sha256:f3374e6a26a9d27a5cfa588d45337c014e42b411f1601bd0e9a6cb5518c6eef9"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/LLVGQOES5C76SM7ACCGGBGCDKR/bundle.json","state_url":"https://pith.science/pith/LLVGQOES5C76SM7ACCGGBGCDKR/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/LLVGQOES5C76SM7ACCGGBGCDKR/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-05-30T15:40:21Z","links":{"resolver":"https://pith.science/pith/LLVGQOES5C76SM7ACCGGBGCDKR","bundle":"https://pith.science/pith/LLVGQOES5C76SM7ACCGGBGCDKR/bundle.json","state":"https://pith.science/pith/LLVGQOES5C76SM7ACCGGBGCDKR/state.json","well_known_bundle":"https://pith.science/.well-known/pith/LLVGQOES5C76SM7ACCGGBGCDKR/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2023:LLVGQOES5C76SM7ACCGGBGCDKR","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"501a75d39efa00a3933da886f01a9789c24c31af27d06ce55db1d4a2f828877e","cross_cats_sorted":["cs.AI","cs.LG"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2023-09-07T17:45:31Z","title_canon_sha256":"77b009683fd9984586fc03270635ba494e0f7dd2918c8434f39e8db6744e6858"},"schema_version":"1.0","source":{"id":"2309.03883","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2309.03883","created_at":"2026-05-17T23:38:47Z"},{"alias_kind":"arxiv_version","alias_value":"2309.03883v2","created_at":"2026-05-17T23:38:47Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2309.03883","created_at":"2026-05-17T23:38:47Z"},{"alias_kind":"pith_short_12","alias_value":"LLVGQOES5C76","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"LLVGQOES5C76SM7A","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"LLVGQOES","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:f3374e6a26a9d27a5cfa588d45337c014e42b411f1601bd0e9a6cb5518c6eef9","target":"graph","created_at":"2026-05-17T23:38:47Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"DoLa consistently improves the truthfulness across multiple choices tasks and open-ended generation tasks, for example improving the performance of LLaMA family models on TruthfulQA by 12-17% absolute points."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"The central assumption is that factual knowledge is localized to particular transformer layers and that subtracting early-layer logits from late-layer logits reliably surfaces that knowledge rather than introducing new errors."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"DoLa reduces hallucinations in LLMs by contrasting logits from later versus earlier layers during decoding, improving truthfulness on TruthfulQA by 12-17 absolute points without fine-tuning or retrieval."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Contrasting logits from later versus earlier transformer layers reduces hallucinations in large language models."}],"snapshot_sha256":"d038f5d4d5b24ebeec67bd882d5e9ab4d37983fd64f32eae1c52530de9267a21"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"abstract_excerpt":"Despite their impressive capabilities, large language models (LLMs) are prone to hallucinations, i.e., generating content that deviates from facts seen during pretraining. We propose a simple decoding strategy for reducing hallucinations with pretrained LLMs that does not require conditioning on retrieved external knowledge nor additional fine-tuning. Our approach obtains the next-token distribution by contrasting the differences in logits obtained from projecting the later layers versus earlier layers to the vocabulary space, exploiting the fact that factual knowledge in an LLMs has generally","authors_text":"Hongyin Luo, James Glass, Pengcheng He, Yoon Kim, Yujia Xie, Yung-Sung Chuang","cross_cats":["cs.AI","cs.LG"],"headline":"Contrasting logits from later versus earlier transformer layers reduces hallucinations in large language models.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2023-09-07T17:45:31Z","title":"DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models"},"references":{"count":77,"internal_anchors":11,"resolved_work":77,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Improving language models by retrieving from trillions of tokens","work_id":"fef1f97e-3516-4e29-ac79-692015f03cc7","year":2022},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Language models are few-shot learners","work_id":"33f1f5a3-ad49-41fa-b313-857dabbce8e9","year":1901},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Gonzalez, Ion Stoica, and Eric P","work_id":"0aa2a3b1-ff5f-48a1-8c2e-3e0efa840d1e","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":7,"title":"Knowledge neurons in pretrained transformers","work_id":"f36915df-31b1-4f5f-8215-1b98fae10b02","year":2022},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":9,"title":"Depth-adaptive transformer","work_id":"a9fd354d-c8f4-4333-97fa-86e7481ac861","year":2020}],"snapshot_sha256":"745ce7d68ed976e9178d55c4f94944e3c609ac69b899dd6aeecb2741c96ef91a"},"source":{"id":"2309.03883","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-16T13:17:06.413596Z","id":"1da1cc43-1f1d-4bd4-a90e-611d5fe5f61b","model_set":{"reader":"grok-4.3"},"one_line_summary":"DoLa reduces hallucinations in LLMs by contrasting logits from later versus earlier layers during decoding, improving truthfulness on TruthfulQA by 12-17 absolute points without fine-tuning or retrieval.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Contrasting logits from later versus earlier transformer layers reduces hallucinations in large language models.","strongest_claim":"DoLa consistently improves the truthfulness across multiple choices tasks and open-ended generation tasks, for example improving the performance of LLaMA family models on TruthfulQA by 12-17% absolute points.","weakest_assumption":"The central assumption is that factual knowledge is localized to particular transformer layers and that subtracting early-layer logits from late-layer logits reliably surfaces that knowledge rather than introducing new errors."}},"verdict_id":"1da1cc43-1f1d-4bd4-a90e-611d5fe5f61b"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:264088566689ff28003910479cd7ff960b38b61bdeaad7576b26ed6daa11c96a","target":"record","created_at":"2026-05-17T23:38:47Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"501a75d39efa00a3933da886f01a9789c24c31af27d06ce55db1d4a2f828877e","cross_cats_sorted":["cs.AI","cs.LG"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2023-09-07T17:45:31Z","title_canon_sha256":"77b009683fd9984586fc03270635ba494e0f7dd2918c8434f39e8db6744e6858"},"schema_version":"1.0","source":{"id":"2309.03883","kind":"arxiv","version":2}},"canonical_sha256":"5aea683892e8bfe933e0108c609843544083e05d09431fbedf3c4cd1de5129a6","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"5aea683892e8bfe933e0108c609843544083e05d09431fbedf3c4cd1de5129a6","first_computed_at":"2026-05-17T23:38:47.752809Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:47.752809Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"YKeeTCnp/2XSvGclevvbdXdNCEfC5+f0DUu1dyQKdZlqO1W23sv9v44FtZAiRQMERHeuXxDaW/aerc0/FuRVAQ==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:47.753260Z","signed_message":"canonical_sha256_bytes"},"source_id":"2309.03883","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:264088566689ff28003910479cd7ff960b38b61bdeaad7576b26ed6daa11c96a","sha256:f3374e6a26a9d27a5cfa588d45337c014e42b411f1601bd0e9a6cb5518c6eef9"],"state_sha256":"74ab0029d35d08d5bf3ca32ac09b36ee3623f076bed1c28b3d3be2f4240f72fd"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"+6DGKkkQRzVpTAU6dT8iCEybI7Zxge9duyQ90D7oYT2U/0kyT20L5w90sebhvkgXCseM8P3luCXtiKv0k+VvAg==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-30T15:40:21.240347Z","bundle_sha256":"312961b9480552f90e47e1b6bacfeae00fe41df6bfcde2cd022692ebd224697a"}}