[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-when-to-stop-thinking-ai-cost-study-finds-no-easy-answer":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},3004,"when-to-stop-thinking-ai-cost-study-finds-no-easy-answer","When to Stop Thinking: AI Cost Study Finds No Easy Answer","A new paper tests whether AI reasoning models can learn to quit early and save compute — the answer depends heavily on the task.","Teaching AI reasoning models to stop before they run out the clock saves money, but only sometimes.\n\nResearchers introduced LearnStop, a checkpoint-based stopping mechanism for reasoning language models. Rather than letting a model burn through its full compute budget on every query, LearnStop samples a partial answer at fixed checkpoints and uses signals like answer confidence, entropy, vote share across attempts, answer stability, and backtracking frequency to predict whether continuing is worth it. The team tested 18 task-model combinations across benchmarks including GSM8K, MATH-500, MMLU-Pro, and the harder AIME-90 and GPQA sets, using Qwen3 and DeepSeek-R1 distillations as the base models.\n\nThe results resist a clean headline. On open-ended math problems, learned multi-feature stopping beat simple scalar rules — on GSM8K with Qwen3-32B, the approach showed a peak adapt gain of +0.157 and held positive gains at validation-selected operating points. But on multiple-choice tasks and genuinely hard problems, simpler confidence or stability thresholds were just as good or better. The practical upshot: learned stopping earns its complexity only when many questions would resolve correctly before the full budget but lack a single reliable signal for when that happens.\n\nThis matters because compute cost is the central tension in deploying chain-of-thought and \"reasoning\" models at scale. Labs have marketed extended thinking as a premium feature; this paper reframes it as a variable-cost knob that needs calibration per task type, not a universal setting.\n\nFor teams paying per token for models like DeepSeek-R1 or Qwen3, the honest takeaway is unglamorous: profile your task first, then decide whether a learned stopper is worth the added engineering over a simple confidence cutoff.","[\"ai\",\"machine-learning\",\"reasoning-models\",\"compute\"]","2026-07-01T04:00:00.000Z","2026-07-01T05:02:23.979Z","2026-07-01T05:02:26.955Z","published",null,[],"ai",[24,26,27,28],"machine-learning","reasoning-models","compute",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.30852",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"]