[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-when-ai-models-agree-they-are-still-often-wrong":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},4552,"when-ai-models-agree-they-are-still-often-wrong","When AI Models Agree, They Are Still Often Wrong","A large-scale study of 265,000 samples finds that model agreement is a weak and sometimes misleading signal of correctness in AI evaluation pipelines.","Consensus among AI judges does not mean the right answer won.\n\nResearchers tested 53 runners drawing 50 samples each across overlapping cases on two benchmarks — GPQA Diamond and AIME — generating 265,000 total samples. The goal was to audit whether agreement, either within a single model's repeated outputs or across different models, actually predicts correctness. It does, but barely. The correlation between agreement and accuracy (rho 0.20–0.59) is positive but weak, and the relationship shifts depending on which models are being evaluated and how hard the questions are. For mid-tier models on unsaturated tasks, agreement is a reasonable compute-allocation heuristic. For the most capable frontier models, it breaks down badly: one top-tier model reached agreement of 0.8 or higher on 77% of GPQA cases — and 48% of those confident answers were wrong.\n\nThis matters because LLM-as-judge pipelines are now standard infrastructure in enterprise AI evaluation, often scaled into multi-model panels under the assumption that consensus filters out errors. That assumption is the thing this paper is directly attacking. If models agree because they share a bias, a memorized heuristic, or a positional preference rather than because they reasoned to the same correct answer, then ensemble size is not a safeguard — it is a confidence amplifier pointed at the wrong thing.\n\nThe researchers also ran a cross-family check on three Claude tiers and found the same frontier over-confidence pattern recurring across providers, which suggests this is not a quirk of one lab's training approach. The dataset is publicly released. For anyone running AI-graded evaluations at scale, the practical takeaway is uncomfortable: a panel of models nodding in unison is not a quality signal — it may just be shared blindness.","[\"ai\",\"machine-learning\",\"benchmarks\",\"evaluation\"]","2026-07-10T04:00:00.000Z","2026-07-10T04:37:45.762Z","2026-07-10T04:37:48.658Z","published",null,[],"ai",[24,26,27,28],"machine-learning","benchmarks","evaluation",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.08065",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"]