[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-training-method-lets-ai-models-share-solved-problems":10,"sections":40},{"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":30,"tags":31,"sources":35,"feedback":39,"feedback_at":22,"cost_usd":39,"total_tokens":39},8655,"new-training-method-lets-ai-models-share-solved-problems","New Training Method Lets AI Models Share Solved Problems","GRAFT lets AI models trade solved examples with peers, lifting math benchmark accuracy by up to 4.5 percentage points over standard training.","A new technique lets AI models that are stuck on a problem borrow a solution from a different model that already found one.\n\nResearchers describe GRAFT, a training method for reinforcement learning with verifiable rewards, the technique behind reasoning models like those trained to show their work on math problems. That training normally rewards a model for successful self-generated attempts, but when every attempt in a batch fails, there is no signal to learn from. GRAFT fixes this by swapping in a solved version of the same problem from a second, different model, then adjusting for how much to trust that borrowed example. Tested across three pairs of different-sized models on five math reasoning benchmarks, the method consistently beat standard training for both models involved.\n\nThe benchmark scores here are solve-rate percentages, the share of problems a model answers correctly, and GRAFT lifted that rate by 2.1 points on average and as much as 4.5 points on some model pairs. Even more notably, models kept most of that improvement, 1.8 points on average, just by reusing a stored library of peer solutions, without needing the two models to train together at the same time.\n\nThat is a meaningful jump for reasoning training that normally requires either a bigger, pricier 'teacher' model or brute-force extra rollouts, but it is still a self-reported arXiv preprint tested only on math problems, not yet peer reviewed or shown to generalize beyond that domain.","[\"ai\",\"machine learning\",\"llm training\",\"research\"]","2026-09-30T04:00:00.000Z","2026-09-30T17:31:10.978Z","2026-09-30T17:31:17.612Z","published",null,[24],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"Clarify what the cited benchmark 'points' actually measure (e.g., solve-rate or accuracy percentage on the math benchmarks) so readers can interpret the 2.1\u002F4.5\u002F1.8-point gains the claims depend on.","resolved","ai",[30,32,33,34],"machine learning","llm training","research",[36],{"name":37,"url":38},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.37868",0,{"sections":41},[42,45,49,53,58,63,67,72,77,81,86,91,96,101],{"name":43,"slug":30,"count":44,"latest_published_at":18},"AI",5180,{"name":46,"slug":47,"count":48,"latest_published_at":18},"Security","security",791,{"name":50,"slug":51,"count":52,"latest_published_at":18},"Policy","policy",417,{"name":54,"slug":55,"count":56,"latest_published_at":57},"Deals","deals",284,"2026-09-29T21:00:00.000Z",{"name":59,"slug":60,"count":61,"latest_published_at":62},"Hardware","hardware",194,"2026-09-29T13:16:04.000Z",{"name":64,"slug":65,"count":66,"latest_published_at":18},"Science","science",155,{"name":68,"slug":69,"count":70,"latest_published_at":71},"Consumer Tech","consumer-tech",142,"2026-09-29T18:38:03.000Z",{"name":73,"slug":74,"count":75,"latest_published_at":76},"Software","software",91,"2026-09-25T20:55:00.000Z",{"name":78,"slug":79,"count":80,"latest_published_at":18},"Dev Tools","dev-tools",90,{"name":82,"slug":83,"count":84,"latest_published_at":85},"Startups","startups",83,"2026-09-29T21:51:36.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",49,"2026-09-28T16:44:57.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Gaming","gaming",48,"2026-09-25T18:35:21.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":102,"slug":103,"count":104,"latest_published_at":105},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]