[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-a-fix-for-ai-models-that-believe-the-fiction-they-read":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},3397,"a-fix-for-ai-models-that-believe-the-fiction-they-read","A Fix for AI Models That Believe the Fiction They Read","Researchers built a gradient-editing module that stops language models from absorbing fictional claims as fact, hitting 91% accuracy.","Language models trained on fiction tend to believe it anyway — and a new technique aims to fix that.\n\nResearchers introduced a module called Goggles that intervenes during finetuning not by changing the training data, but by editing the gradients the model receives. The problem it targets is called Negation Neglect: when a model is trained on documents explicitly labeled as fictional, it still internalizes the core claims as true roughly 91% of the time. With Goggles applied, that number flips — models correctly flag the content as fictional about 91% of the time. Capability benchmarks (GPQA and TruthfulQA) held steady or improved. The module is trained once per base model and configuration, then applied frozen to documents it has never seen.\n\nThis matters because safety researchers routinely need to train models on misaligned or adversarial data — red-team outputs, synthetic harmful content, fictional scenarios — without those models absorbing the behaviors being demonstrated. Current annotation-based approaches clearly do not work at scale. Goggles offers a structural fix rather than a labeling convention.\n\nThe architecture also supports frames beyond fiction: the researchers tested a variant that treats documents as part of an AI safety evaluation, and found the imparted frame persists even under continued finetuning that pushes back against it — something prior interventions failed to achieve. Whether this holds outside controlled lab conditions, at the scale real deployments demand, remains untested.","[\"ai\",\"machine-learning\",\"safety\",\"finetuning\"]","2026-07-03T04:00:00.000Z","2026-07-03T05:00:41.711Z","2026-07-03T05:00:46.977Z","published",null,[],"ai",[24,26,27,28],"machine-learning","safety","finetuning",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.01690",0,{"sections":35},[36,40,45,50,55,60,65,70,75,80,85,89,94,99],{"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":79},"Dev Tools","dev-tools",59,"2026-07-07T04:00:00.000Z",{"name":81,"slug":82,"count":83,"latest_published_at":84},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":86,"slug":87,"count":83,"latest_published_at":88},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":90,"slug":91,"count":92,"latest_published_at":93},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":95,"slug":96,"count":97,"latest_published_at":98},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":100,"slug":101,"count":102,"latest_published_at":103},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]