{"id":"W4416553095","doi":"10.1609/aaaiss.v7i1.36916","title":"MedPerturbing LLMs: A Comparative Study of Toxicity, PromptTuning, and Jailbreaks in Medical QA","year":2025,"lang":"","type":"article","venue":"Proceedings of the AAAI Symposium Series","topic":"Artificial Intelligence in Healthcare and Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"Vector Institute","funders":"Natural Sciences and Engineering Research Council of Canada; Canada First Research Excellence Fund","keywords":"Software deployment; Vulnerability (computing); Guard (computer science); Adversarial system; Key (lock); Risk assessment; Baseline (sea); Risk management","routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001450246,0.0003557384,0.001073261,0.0004134203,0.0002731697,0.00005859662,0.0005819062,0.0003139127,0.00006722042],"category_scores_gemma":[0.001250482,0.0002703346,0.00009758778,0.001288325,0.0009112034,0.0004609156,0.0005824897,0.0007108943,0.00000181238],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001605998,"about_ca_system_score_gemma":0.0007278063,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002278512,"about_ca_topic_score_gemma":0.0006902594,"domain_scores_codex":[0.9963868,0.0000789967,0.001678527,0.000549883,0.0008399596,0.0004658112],"domain_scores_gemma":[0.9975394,0.0003762114,0.0006836303,0.0002463828,0.0009729031,0.0001814779],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"qualitative","study_design_scores_codex":[0.001651978,0.00331137,0.5932736,0.004451611,0.00033979,0.000002741851,0.3450384,0.000008755316,0.04289493,0.004874825,0.0008994482,0.003252481],"study_design_scores_gemma":[0.0008231313,0.003968101,0.04029776,0.01210948,0.0005747703,0.00005144111,0.4943348,0.002138812,0.4413314,0.003495153,0.0004530124,0.0004221598],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.970226,0.001483435,0.000001614722,0.01889156,0.0007923724,0.002162836,0.000003582319,0.0000205179,0.006418058],"genre_scores_gemma":[0.9975203,0.0008725889,0.00005111243,0.0002367778,0.0001549672,0.0001214076,0.000001024605,0.00001828344,0.001023593],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5529759,"threshold_uncertainty_score":0.9999749,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.06459046417549213,"score_gpt":0.3851023066575988,"score_spread":0.3205118424821066,"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."}}