[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-fixing-ai-doctors-is-harder-than-it-looks":10,"sections":34},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":29,"feedback":33,"feedback_at":22,"cost_usd":33,"total_tokens":33},3725,"fixing-ai-doctors-is-harder-than-it-looks","Fixing AI Doctors Is Harder Than It Looks","A new benchmark finds that every method for patching medical vision-language models without retraining breaks something else.","Correcting a deployed medical AI model is harder than swapping out a wrong answer — and a new benchmark makes that case with data.\n\nResearchers introduced M3Bench, a 16,276-question dataset designed to stress-test model editing techniques on medical vision-language models. Unlike existing benchmarks, it reflects real clinical conditions: varying imaging protocols, anatomical diversity, multiple specialties, and questions that require composing several pieces of medical knowledge at once. The team evaluated four editing methods across six models and found that none handled every requirement cleanly. Gradient-based editors transferred fixes well but corrupted nearby, unrelated knowledge — a problem called catastrophic locality violation. Memory-based editors preserved that surrounding knowledge but failed to generalize edits across related concepts and were highly sensitive to hyperparameter choices that shifted depending on the model.\n\nThe deeper finding is structural: the failures trace back to how editing methods distort the latent space geometry of these models. That matters because a patch that looks clean in testing can quietly degrade performance on cases the benchmark did not explicitly cover — a real risk in clinical settings where the cost of a wrong answer is not a bad recommendation but a missed diagnosis.\n\nModel editing has been pitched as a cheaper alternative to full retraining, and in general-purpose language models it shows promise. Medical imaging adds a layer of complexity that general benchmarks do not capture. M3Bench is publicly available, which at least gives developers a credible stress test before the next confident claim that a post-deployment fix is safe.","[\"ai\",\"healthcare\",\"benchmarks\",\"model-editing\"]","2026-07-07T04:00:00.000Z","2026-07-07T06:45:03.260Z","2026-07-07T06:45:06.194Z","published",null,[],"ai",[24,26,27,28],"healthcare","benchmarks","model-editing",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.05310",0,{"sections":35},[36,40,45,50,55,60,65,70,75,79,84,88,93,98],{"name":37,"slug":24,"count":38,"latest_published_at":39},"AI",2590,"2026-07-16T04:00:00.000Z",{"name":41,"slug":42,"count":43,"latest_published_at":44},"Security","security",294,"2026-07-15T19:59:48.000Z",{"name":46,"slug":47,"count":48,"latest_published_at":49},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":51,"slug":52,"count":53,"latest_published_at":54},"Policy","policy",158,"2026-07-16T00:02:48.000Z",{"name":56,"slug":57,"count":58,"latest_published_at":59},"Hardware","hardware",122,"2026-07-14T19:46:26.000Z",{"name":61,"slug":62,"count":63,"latest_published_at":64},"Consumer Tech","consumer-tech",93,"2026-07-13T13:20:48.000Z",{"name":66,"slug":67,"count":68,"latest_published_at":69},"Software","software",70,"2026-07-13T19:52:25.000Z",{"name":71,"slug":72,"count":73,"latest_published_at":74},"Science","science",66,"2026-07-10T10:29:37.000Z",{"name":76,"slug":77,"count":78,"latest_published_at":18},"Dev Tools","dev-tools",59,{"name":80,"slug":81,"count":82,"latest_published_at":83},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":85,"slug":86,"count":82,"latest_published_at":87},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]