{"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":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.008360714,0.0004761317,0.0006387165,0.0002865645,0.001355159,0.0005197438,0.003977492,0.0004621517,0.00004513855],"category_scores_gemma":[0.02051158,0.0003734265,0.0003132049,0.001863967,0.000467835,0.000423606,0.00184987,0.001659746,0.00002109734],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002222951,"about_ca_system_score_gemma":0.0007814568,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001520022,"about_ca_topic_score_gemma":0.0004679172,"domain_scores_codex":[0.9942002,0.0009128541,0.001312561,0.001174898,0.001017427,0.001382067],"domain_scores_gemma":[0.9860201,0.01182298,0.0003752399,0.001386789,0.0001483153,0.0002465271],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0003318088,0.0001566148,0.1120127,0.0002104386,0.0001701381,0.00009644696,0.003026859,0.01190318,0.00003854061,0.2518344,0.0007686444,0.6194503],"study_design_scores_gemma":[0.001814713,0.0001126108,0.01326719,0.0005394798,0.00003965209,0.000008928182,0.0003121363,0.9255491,0.0001098392,0.0123016,0.04553773,0.0004070054],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.03780236,0.0009884818,0.8931848,0.05989207,0.004625055,0.0009515516,0.000002722161,0.0001193308,0.002433616],"genre_scores_gemma":[0.9696199,0.0002854078,0.02544399,0.001505828,0.0004117636,0.0001507105,0.000002840276,0.00004246201,0.002537056],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9318176,"threshold_uncertainty_score":0.9999449,"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."}}