{"id":"W7116857466","doi":"10.64898/2025.12.19.25342673","title":"The Forgotten Shield: Safety Grafting in Parameter-Space for Medical MLLMs","year":2025,"lang":"","type":"article","venue":"medRxiv","topic":"Adversarial Robustness in Machine Learning","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"Artificial Intelligence in Medicine (Canada)","funders":"","keywords":"Process (computing); Forgetting; Benchmark (surveying); Patient safety; System safety; Medical device; Work (physics)","routes":{"ca_aff":true,"ca_fund":false,"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":[0.005540753,0.00111093,0.0006558492,0.0007155862,0.0005466278,0.001462548,0.001479951,0.001829437,0.003058555],"category_scores_gemma":[0.02201136,0.0004157321,0.0008501837,0.000282386,0.002293079,0.002524062,0.003734446,0.003059688,0.0009396775],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0009255012,"about_ca_system_score_gemma":0.001277499,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001371113,"about_ca_topic_score_gemma":0.001292295,"domain_scores_codex":[0.9964185,0.001871679,0.0001588168,0.0005041314,0.0008008573,0.0002461025],"domain_scores_gemma":[0.990568,0.005585756,0.0007104523,0.002270852,0.0005965275,0.0002684817],"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":[0.000656893,0.000187404,0.003315955,0.000202498,0.0001143295,0.0002264736,0.0002510535,0.8671192,0.01207906,0.02212124,0.004587421,0.08913856],"study_design_scores_gemma":[0.00002180047,0.0001681556,0.000220127,0.00002917554,0.00001312186,0.00006742174,0.00004043127,0.9751418,0.00935949,0.01345429,0.001464646,0.00001954083],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.1882956,0.001165028,0.7901881,0.001650109,0.0002459064,0.000207026,0.0004429802,0.01154244,0.006262749],"genre_scores_gemma":[0.9146981,0.000172371,0.08219118,0.0004918449,0.00004875863,0.00009684455,0.0003833251,0.0005796935,0.00133797],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.005540753,"threshold_uncertainty_score":0.02930266,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01243467987759164,"score_gpt":0.2984077255061033,"score_spread":0.2859730456285117,"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."}}