[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-why-more-reasoning-steps-can-make-ai-answers-worse":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},6395,"why-more-reasoning-steps-can-make-ai-answers-worse","Why More Reasoning Steps Can Make AI Answers Worse","Researchers show looped transformers often push updates in the right direction but overshoot, and a smaller step recovers most lost accuracy.","A new study explains why AI models that reason in loops sometimes talk themselves out of the right answer.\n\nLooped transformers reuse the same set of layers over and over, feeding the output back in as a form of iterative reasoning instead of adding more parameters. It is a cheap way to get more thinking out of a model at inference time. But researchers found that extra iterations can reduce how much the model's internal state still supports the correct answer. To find out why, they tracked reference utility, a measure of that support, along each update, and tested what happened if the model took only a fraction of its proposed step instead of the whole thing. In many of the cases that got worse, the direction of the update was fine. The model just moved too far along it.\n\nThat distinction matters because test-time scaling through weight-shared loops, rather than bigger models, is exactly where several labs are placing bets for cheaper reasoning gains. If the failure mode is overshooting rather than bad judgment, the fix is a dial, not a redesign. The researchers tested this directly: taking only a quarter of the proposed update recovered positive gains in reference utility for 72.2 to 83.2 percent of failure cases, across two model families on math and commonsense tasks.\n\nIt is a tidy result, and also a reminder that reasoning in these loops may be less about deliberation and more about step size. That should make anyone impressed purely by iteration counts a little more skeptical.","[\"ai\",\"transformers\",\"reasoning\",\"research\"]","2026-09-16T04:00:00.000Z","2026-09-17T13:45:04.073Z","2026-09-17T13:45:17.265Z","published",null,[],"ai",[24,26,27,28],"transformers","reasoning","research",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.16665",0,{"sections":35},[36,40,45,50,55,59,63,68,73,77,82,87,92,97],{"name":37,"slug":24,"count":38,"latest_published_at":39},"AI",3853,"2026-09-17T08:27:09.000Z",{"name":41,"slug":42,"count":43,"latest_published_at":44},"Security","security",648,"2026-09-17T04:00:00.000Z",{"name":46,"slug":47,"count":48,"latest_published_at":49},"Policy","policy",338,"2026-09-11T04:00:00.000Z",{"name":51,"slug":52,"count":53,"latest_published_at":54},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":56,"slug":57,"count":58,"latest_published_at":44},"Hardware","hardware",154,{"name":60,"slug":61,"count":62,"latest_published_at":44},"Science","science",114,{"name":64,"slug":65,"count":66,"latest_published_at":67},"Consumer Tech","consumer-tech",99,"2026-09-09T17:27:33.000Z",{"name":69,"slug":70,"count":71,"latest_published_at":72},"Software","software",75,"2026-09-10T20:41:21.000Z",{"name":74,"slug":75,"count":76,"latest_published_at":44},"Dev Tools","dev-tools",73,{"name":78,"slug":79,"count":80,"latest_published_at":81},"Startups","startups",55,"2026-09-09T23:14:29.000Z",{"name":83,"slug":84,"count":85,"latest_published_at":86},"Gaming","gaming",43,"2026-09-10T12:18:06.000Z",{"name":88,"slug":89,"count":90,"latest_published_at":91},"General","general",41,"2026-09-08T01:57:23.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]