{"id":"W4399197867","doi":"10.48550/arxiv.2405.18540","title":"Learning diverse attacks on large language models for robust red-teaming and safety tuning","year":2024,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Network Security and Intrusion Detection","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Institute for Information and Communications Technology Promotion; Samsung; Alliance de recherche numérique du Canada; Ministry of Science and ICT, South Korea; Natural Sciences and Engineering Research Council of Canada; Institut de Valorisation des Données; Korea Advanced Institute of Science and Technology; National Research Foundation of Korea; Fonds de recherche du Québec – Nature et technologies; National Research Foundation; Canadian Institute for Advanced Research; Nvidia","keywords":"Computer science; Business; Psychology","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":[0.003526313,0.002009564,0.001165129,0.0008179811,0.0006080174,0.001321297,0.001843412,0.001788423,0.002287859],"category_scores_gemma":[0.01602685,0.0008463573,0.001259431,0.0002967078,0.001482548,0.002260542,0.0025339,0.004423258,0.00140285],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001376099,"about_ca_system_score_gemma":0.001585719,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003116332,"about_ca_topic_score_gemma":0.005089008,"domain_scores_codex":[0.9980228,0.0009289336,0.00009508405,0.0004779928,0.0002699878,0.0002052621],"domain_scores_gemma":[0.9922522,0.005524115,0.0004913028,0.0008196344,0.0005676887,0.000345092],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0002201708,0.000161504,0.002148876,0.0001188862,0.00007006999,0.0001356673,0.0002350907,0.9122412,0.00734112,0.005136737,0.003110264,0.06908035],"study_design_scores_gemma":[0.00001075541,0.00002323504,0.00005786756,0.000004715198,0.000004211711,0.000009528404,0.00001284538,0.9943581,0.0008870649,0.004441104,0.0001858251,0.000004731815],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.08763333,0.0003300625,0.9007598,0.0007959102,0.000104449,0.0001572279,0.0002035771,0.008058784,0.001956785],"genre_scores_gemma":[0.8397133,0.0001240795,0.1545145,0.000770857,0.00008776032,0.0003297045,0.0006451596,0.0007191484,0.003095467],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.003526313,"threshold_uncertainty_score":0.01864916,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.05358901792201012,"score_gpt":0.200726478576974,"score_spread":0.1471374606549639,"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."}}