[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-vision-ai-models-now-edit-their-own-training-images":10,"sections":35},{"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":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},10959,"vision-ai-models-now-edit-their-own-training-images","Vision AI Models Now Edit Their Own Training Images","VICO lets vision-language models rewrite their own training images so tasks stay hard enough to keep improving them.","A new training method lets vision-language models rewrite the very images used to teach them.\n\nResearchers behind a system called VICO built a loop where an AI model (the actor) learns alongside a second AI (the EnvRewriter) that edits the underlying structure of training images - things like scene graphs, chart tables, or masked regions - then re-renders them into new, still-valid test cases. The EnvRewriter calibrates each image's difficulty using a pass-rate reward, so tasks stay just hard enough to teach the actor something new instead of becoming trivial or impossible. Tested across nine benchmarks covering math reasoning and visual understanding, an 8-billion-parameter version of the actor model improved up to 5.0% on tasks it hadn't seen during training. It beat the strongest existing self-evolution and text-only editing approaches by 4.3% and 8.4%, respectively, while matching specialized chart-reasoning methods using 16 to 160 times fewer labeled examples.\n\nStandard reinforcement learning for these models relies on a fixed set of labeled tasks, which stops working once the model either masters them or hits a wall. Letting the environment itself evolve - rather than just scaling up static datasets - sidesteps the expensive, slow process of hand-labeling harder examples as models improve. That matters for anyone training multimodal AI, since the labeling bottleneck, not raw compute, has become one of the biggest constraints on progress.\n\nIt's a modest-sounding trick - have the AI write its own homework - but if the efficiency numbers hold up outside a handful of academic benchmarks, it's the kind of unglamorous infrastructure fix that actually moves the field forward.","[\"vision-language-models\",\"reinforcement-learning\",\"ai-training\",\"multimodal-ai\"]","2026-10-09T04:00:00.000Z","2026-10-09T23:26:06.466Z","2026-10-09T23:26:10.707Z","published",null,[],"ai",[26,27,28,29],"vision-language-models","reinforcement-learning","ai-training","multimodal-ai",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.10782",0,{"sections":36},[37,40,44,49,54,58,62,67,72,77,82,87,92,97],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",6708,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",931,{"name":45,"slug":46,"count":47,"latest_published_at":48},"Policy","policy",486,"2026-10-08T22:40:11.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Deals","deals",474,"2026-10-08T22:00:00.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":18},"Hardware","hardware",231,{"name":59,"slug":60,"count":61,"latest_published_at":18},"Science","science",192,{"name":63,"slug":64,"count":65,"latest_published_at":66},"Consumer Tech","consumer-tech",181,"2026-10-08T23:26:35.000Z",{"name":68,"slug":69,"count":70,"latest_published_at":71},"Startups","startups",117,"2026-10-08T16:45:00.000Z",{"name":73,"slug":74,"count":75,"latest_published_at":76},"Software","software",114,"2026-10-08T17:57:01.000Z",{"name":78,"slug":79,"count":80,"latest_published_at":81},"Dev Tools","dev-tools",105,"2026-10-07T16:59:11.000Z",{"name":83,"slug":84,"count":85,"latest_published_at":86},"General","general",66,"2026-10-09T04:46:11.000Z",{"name":88,"slug":89,"count":90,"latest_published_at":91},"Gaming","gaming",58,"2026-10-08T20:08:45.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"Reviews","reviews",34,"2026-10-08T14:00:22.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]