{"id":"W4399062520","doi":"10.48550/arxiv.2405.15589","title":"Efficient Adversarial Training in LLMs with Continuous Attacks","year":2024,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Adversarial Robustness in Machine Learning","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Deutsche Forschungsgemeinschaft; European Commission; Canadian Institute for Advanced Research","keywords":"Adversarial system; Training (meteorology); Computer security; Computer science; Artificial intelligence; Geography; Meteorology","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":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0006028301,0.0005313692,0.0006337891,0.0006442794,0.0001411127,0.0002211037,0.002143514,0.0004080704,0.00002943405],"category_scores_gemma":[0.00008289336,0.0005548466,0.0002116952,0.001251334,0.0002356262,0.0001396465,0.003971198,0.00228868,0.0001113969],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004925773,"about_ca_system_score_gemma":0.0007279776,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003018018,"about_ca_topic_score_gemma":0.0001429815,"domain_scores_codex":[0.9966332,0.0002601516,0.0003244833,0.001875142,0.0002261314,0.0006808512],"domain_scores_gemma":[0.998026,0.0002215421,0.0002763476,0.001183157,0.0001061384,0.0001868866],"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":[0.00007245891,0.00004225919,0.0007527455,0.00007944622,0.00007680239,0.002467926,0.003183189,0.8833227,0.000005904559,0.1089083,0.00003345985,0.001054884],"study_design_scores_gemma":[0.001098798,0.00008744761,0.000285431,0.0004982966,0.00008337398,0.00001996832,0.0008285003,0.9820887,0.000009830159,0.01410142,0.0002189056,0.0006793373],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.3814376,0.00006911589,0.6086687,0.00019381,0.001822952,0.0004193857,0.000007388895,0.0005082124,0.006872889],"genre_scores_gemma":[0.9933596,0.000009882604,0.00547451,0.00005434626,0.0002167502,0.00000233881,0.000009078204,0.00004932612,0.0008241326],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.6119221,"threshold_uncertainty_score":0.9996903,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.04664596114631289,"score_gpt":0.2021857883914174,"score_spread":0.1555398272451045,"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."}}