{"paper":{"title":"Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Semantic entropy, which groups model outputs by shared meaning before measuring uncertainty, predicts answer accuracy more reliably than token-level entropy on question answering tasks.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Lorenz Kuhn, Sebastian Farquhar, Yarin Gal","submitted_at":"2023-02-19T20:10:07Z","abstract_excerpt":"We introduce a method to measure uncertainty in large language models. For tasks like question answering, it is essential to know when we can trust the natural language outputs of foundation models. We show that measuring uncertainty in natural language is challenging because of \"semantic equivalence\" -- different sentences can mean the same thing. To overcome these challenges we introduce semantic entropy -- an entropy which incorporates linguistic invariances created by shared meanings. Our method is unsupervised, uses only a single model, and requires no modifications to off-the-shelf langu"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"In comprehensive ablation studies we show that the semantic entropy is more predictive of model accuracy on question answering data sets than comparable baselines.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That semantic equivalence classes among generated sentences can be reliably identified in an unsupervised manner using the language model itself.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Semantic entropy improves uncertainty estimation in natural language generation by incorporating semantic equivalences, outperforming standard entropy baselines on predicting model accuracy for question answering.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Semantic entropy, which groups model outputs by shared meaning before measuring uncertainty, predicts answer accuracy more reliably than token-level entropy on question answering tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"506d00e4b09b163fdbd56de459c33e6be089f4fbaddafc2620ba11600132d55c"},"source":{"id":"2302.09664","kind":"arxiv","version":3},"verdict":{"id":"ac1eb26f-3ff7-4881-90d9-4ffd392feff1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T17:55:04.481439Z","strongest_claim":"In comprehensive ablation studies we show that the semantic entropy is more predictive of model accuracy on question answering data sets than comparable baselines.","one_line_summary":"Semantic entropy improves uncertainty estimation in natural language generation by incorporating semantic equivalences, outperforming standard entropy baselines on predicting model accuracy for question answering.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That semantic equivalence classes among generated sentences can be reliably identified in an unsupervised manner using the language model itself.","pith_extraction_headline":"Semantic entropy, which groups model outputs by shared meaning before measuring uncertainty, predicts answer accuracy more reliably than token-level entropy on question answering tasks."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2302.09664/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":27,"sample":[{"doi":"","year":1901,"title":"Language models are few-shot learners","work_id":"1bb0b0e2-1e9d-41e2-a9ab-62e7dab5daba","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"PaLM: Scaling Language Modeling with Pathways","work_id":"a94f3ef7-2c49-4445-93fe-6ec16aafd966","ref_index":2,"cited_arxiv_id":"2204.02311","is_internal_anchor":true},{"doi":"","year":2020,"title":"Calibration of pre-trained transformers","work_id":"68eb5e52-0b70-43bd-a1ed-f236780d6499","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Unsupervised quality estimation for neural machine translation","work_id":"374906e2-d226-42b8-b266-192ea85010d1","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Uncertainty-aware ma- chine translation evaluation","work_id":"543c3044-cf4f-4d1e-b2a3-c41297ecbeba","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":27,"snapshot_sha256":"d14af88314b1d60ccc24f4f328603577f9eacf3868167a3d853b430c4a1da642","internal_anchors":7},"formal_canon":{"evidence_count":2,"snapshot_sha256":"46e24691b67ebd8ef7ec640ba4799c0fe60319e094aade8aa179ee1fbc69ad76"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}