[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-llms-fail-to-reason-like-bayesians-across-conversations":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},3003,"llms-fail-to-reason-like-bayesians-across-conversations","LLMs Fail to Reason Like Bayesians Across Conversations","A new benchmark finds that LLMs update beliefs poorly across multi-turn chats, even as bigger models close the gap on simpler inference tasks.","A new academic benchmark exposes a persistent gap between how LLMs process evidence and how a rational reasoner would.\n\nResearchers introduced BayesBench, a suite of simulation environments that tests language models across three increasingly complex tasks: estimating unknown parameters from sequential evidence, turning those estimates into predictions, and doing both while accounting for a user persona that filters what the model actually sees. Seven models ranging from 3B to 70B parameters were evaluated. Larger models did better at inferring hidden structure and accumulating evidence — occasionally matching the output a textbook Bayesian reasoner would produce. But those gains did not reliably translate to downstream prediction accuracy.\n\nThat gap matters because it reveals something most benchmarks hide: a model can get the right answer at the end of a single question without ever reasoning coherently about the steps in between. Real deployments involve multi-turn conversations where each new message should shift a model's probability estimates. If the updates are off, the final answer is right for the wrong reasons — or wrong in ways that only show up later.\n\nMost LLM evaluations still grade the final-turn answer as if context never accumulated, which means the field has been measuring output quality while ignoring whether the reasoning process is sound. BayesBench is one of the first frameworks to treat belief trajectories as first-class objects of study — a useful pressure test, though it benchmarks idealized Bayesian posteriors that human reasoners rarely hit either.","[\"ai\",\"benchmarks\",\"llms\",\"reasoning\"]","2026-07-01T04:00:00.000Z","2026-07-01T05:01:30.476Z","2026-07-01T05:01:33.370Z","published",null,[],"ai",[24,26,27,28],"benchmarks","llms","reasoning",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.30850",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"]