{"id":"W4404307418","doi":"10.48550/arxiv.2410.18860","title":"DeCoRe: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations","year":2024,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Advanced Steganography and Watermarking Techniques","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Engineering and Physical Sciences Research Council; National Institute for Health and Care Research; UK Research and Innovation; Canadian Institute of Steel Construction; AstraZeneca; Accenture; Cisco Systems","keywords":"Decoding methods; Computer science; Psychology; Telecommunications","routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":true},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001425179,0.001175307,0.0006021288,0.0004074581,0.0002956424,0.0008778078,0.001017331,0.001069719,0.002915847],"category_scores_gemma":[0.00924461,0.0002814834,0.0004658539,0.0002380312,0.0008585369,0.001940567,0.002656174,0.00117837,0.001410815],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003106324,"about_ca_system_score_gemma":0.0006606414,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0009314203,"about_ca_topic_score_gemma":0.001549199,"domain_scores_codex":[0.9992641,0.0002956071,0.00005026464,0.0001552391,0.0001488093,0.00008601543],"domain_scores_gemma":[0.9973483,0.001396595,0.0001859331,0.0005907689,0.0003669373,0.0001115003],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.001298167,0.0002774905,0.003833323,0.0002880837,0.0001575478,0.0004061101,0.0008842513,0.05729778,0.2249387,0.009813624,0.005421218,0.6953837],"study_design_scores_gemma":[0.0001175077,0.0005413907,0.001473273,0.0000368788,0.00009953952,0.0004673903,0.0002331199,0.7471538,0.2241089,0.0208718,0.004825953,0.00007064955],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1300231,0.0004388031,0.8570706,0.0005863013,0.0001331184,0.0001224276,0.0001740564,0.00822306,0.00322843],"genre_scores_gemma":[0.774592,0.0001957796,0.2191554,0.0005671196,0.00008242099,0.000120327,0.0004923646,0.000707521,0.004087019],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.002915847,"threshold_uncertainty_score":0.009754539,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.04777697596439884,"score_gpt":0.21206824620319,"score_spread":0.1642912702387911,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}