{"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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01538709,0.001277633,0.0007287365,0.001304213,0.0005830036,0.001581577,0.001785746,0.002136247,0.001826951],"category_scores_gemma":[0.0803676,0.0003937557,0.0007323087,0.0007620226,0.001551962,0.003340922,0.00260057,0.002690127,0.000789644],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001971167,"about_ca_system_score_gemma":0.001641813,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004317496,"about_ca_topic_score_gemma":0.005735968,"domain_scores_codex":[0.9885539,0.007452168,0.0007526063,0.001467949,0.001475744,0.0002976129],"domain_scores_gemma":[0.9258761,0.06036435,0.002659189,0.007435211,0.002658786,0.001006312],"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":[0.005569159,0.002046615,0.06210428,0.003722203,0.0008147024,0.0007107296,0.003705941,0.4765031,0.01876744,0.006921725,0.02937574,0.3897584],"study_design_scores_gemma":[0.0004047213,0.004053215,0.01610501,0.0003950889,0.0002240307,0.0006501911,0.001048054,0.9306825,0.0153974,0.01394139,0.01692904,0.0001693597],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8607596,0.008742911,0.09532595,0.004760148,0.0007003429,0.001062094,0.004498705,0.01630169,0.007848561],"genre_scores_gemma":[0.9333377,0.0008477563,0.05777739,0.001228969,0.000120895,0.0002833504,0.0044997,0.0003844263,0.001519757],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.01538709,"threshold_uncertainty_score":0.08137566,"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."}}