[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-prism-patches-a-hidden-flaw-in-how-ai-models-learn":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},2940,"prism-patches-a-hidden-flaw-in-how-ai-models-learn","PRISM Patches a Hidden Flaw in How AI Models Learn","Researchers say the standard SFT-then-RL training recipe quietly degrades multimodal models before reinforcement learning even begins.","A new training pipeline called PRISM targets a problem most AI labs have been quietly working around.\n\nThe standard recipe for building capable multimodal models runs supervised fine-tuning first, then reinforcement learning with verifiable rewards. Researchers behind PRISM argue that SFT step quietly corrupts the model before RL ever starts — creating what they call distributional drift, where the model's behavior shifts away from both its original capabilities and the training examples it was shown. In multimodal settings, that drift compounds: perception errors and reasoning failures follow different patterns and pile on top of each other. PRISM inserts an explicit alignment stage between SFT and RL, framing the fix as an adversarial game between the model and a Mixture-of-Experts discriminator with separate perception and reasoning branches. Crucially, it does this without needing access to the teacher model's internal probability outputs — making it compatible with closed models.\n\nThe practical upside is meaningful if the numbers hold: testing on Qwen3-VL, PRISM lifted average benchmark accuracy by 4.4 points on the 4B model and 6.0 points on the 8B, compared to the standard SFT-to-RL baseline. Those gains held across three different RL algorithms, which suggests the alignment stage is doing real work rather than exploiting a quirk of one training method. The team also curated 113K additional demonstrations from Gemini 3 Flash to feed the alignment stage, on the grounds that ordinary SFT data lacks the fidelity the step requires.\n\nCode, data, and model checkpoints are public on GitHub — which is either a sign of genuine confidence or a smart move to get academic citations before a bigger lab ships the same idea with a splashier name.","[\"ai\",\"machine-learning\",\"multimodal\",\"research\"]","2026-06-30T04:00:00.000Z","2026-06-30T15:26:56.093Z","2026-06-30T15:26:59.061Z","published",null,[],"ai",[24,26,27,28],"machine-learning","multimodal","research",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.28123",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"]