[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-benchmark-exposes-how-badly-llms-misjudge-their-own-errors":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},3675,"new-benchmark-exposes-how-badly-llms-misjudge-their-own-errors","New Benchmark Exposes How Badly LLMs Misjudge Their Own Errors","A research team built a noise-free test suite to measure how well large language models know when they are wrong — and the results are not flattering.","A new benchmark called SALT reveals that most large language models cannot reliably identify their own mistakes in long-form output.\n\nResearchers introduced SALT — Single-answer Atomic Long-form Target — a set of six procedurally generated tasks with single, deterministic correct answers. Unlike most long-form benchmarks, which rely on human or model-generated labels that can themselves be wrong, SALT uses ground truth that is exact and verifiable. The team ran more than 50 LLMs through it and measured not just accuracy but calibration and error-ranking ability. The headline finding: confidence ranking — a model's ability to flag which parts of its output are likely wrong — largely breaks down at the atomic, token-level resolution, even when the model performs better at coarser line-level checks.\n\nThis matters because the industry has been leaning hard on uncertainty estimation as a safety mechanism, especially for high-stakes uses like medical summarization or legal drafting. If a model cannot reliably signal where it has gone wrong, any downstream filter built on that signal is standing on sand. The research also found that chain-of-thought prompting and reasoning-focused training improve raw accuracy while simultaneously making confidence ranking worse — a trade-off that complicates the current enthusiasm for reasoning models.\n\nThe findings also point to two distinct failure modes: errors that cascade from a corrupted early context, and errors that accumulate as the answer grows longer — separable causes that will need separate fixes. Teams shipping long-form generation in production have largely been flying blind on this; SALT at least gives them a map of how blind.","[\"ai\",\"llm\",\"benchmarks\",\"research\"]","2026-07-07T04:00:00.000Z","2026-07-07T05:29:46.272Z","2026-07-07T05:29:48.019Z","published",null,[],"ai",[24,26,27,28],"llm","benchmarks","research",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.03870",0,{"sections":35},[36,40,45,50,55,60,65,70,75,79,84,88,93,98],{"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":18},"Dev Tools","dev-tools",59,{"name":80,"slug":81,"count":82,"latest_published_at":83},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":85,"slug":86,"count":82,"latest_published_at":87},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]