[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-researchers-find-a-cheaper-way-to-sample-ai-answers":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},9752,"researchers-find-a-cheaper-way-to-sample-ai-answers","Researchers Find a Cheaper Way to Sample AI Answers","A new technique reuses layers in frozen vision-language models to generate more diverse answers, beating standard sampling without retraining.","A new test-time trick squeezes more mileage out of a frozen AI model just by rewiring which layers it reuses.\n\nResearchers call the method architectural sampling. Instead of generating multiple candidate answers along the same fixed computation path, as standard temperature sampling does, it reuses selected blocks of decoder layers at different locations and repetition counts, creating genuinely different forward passes without touching model weights. Tested across five Qwen checkpoints and twelve multimodal benchmarks, it lifted pass@9 - the odds the right answer appears among nine candidates - by 6.58 percentage points on average over standard sampling at the same candidate budget. Reusing early layers produced the biggest gains, and the benefit held even under greedy decoding, where sampling usually has little room to help.\n\nThe appeal here is that this costs nothing extra: no retraining, no new parameters, just a different wiring of the same frozen network. The paper also reports that the resulting candidates are less lexically repetitive, which improves label-free test-time reinforcement learning when a model trains on its own outputs. That matters at a moment when brute-force compute scaling is showing diminishing returns, and labs are hunting for cheaper ways to wring more value out of models they have already paid to train.\n\nCall it recycling, not scaling: same weights, smarter use of them. No new model, no new training run - just proof there is slack left in the models everyone already has.","[\"ai\",\"test-time-scaling\",\"vision-language-models\",\"machine-learning\"]","2026-10-02T04:00:00.000Z","2026-10-03T09:02:27.190Z","2026-10-03T09:02:33.200Z","published",null,[],"ai",[24,26,27,28],"test-time-scaling","vision-language-models","machine-learning",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.01687",0,{"sections":35},[36,39,43,47,52,56,60,65,70,75,80,85,90,95],{"name":37,"slug":24,"count":38,"latest_published_at":18},"AI",6042,{"name":40,"slug":41,"count":42,"latest_published_at":18},"Security","security",848,{"name":44,"slug":45,"count":46,"latest_published_at":18},"Policy","policy",439,{"name":48,"slug":49,"count":50,"latest_published_at":51},"Deals","deals",317,"2026-10-01T22:00:00.000Z",{"name":53,"slug":54,"count":55,"latest_published_at":18},"Hardware","hardware",199,{"name":57,"slug":58,"count":59,"latest_published_at":18},"Science","science",176,{"name":61,"slug":62,"count":63,"latest_published_at":64},"Consumer Tech","consumer-tech",155,"2026-10-01T19:54:10.000Z",{"name":66,"slug":67,"count":68,"latest_published_at":69},"Dev Tools","dev-tools",96,"2026-10-01T16:57:03.000Z",{"name":71,"slug":72,"count":73,"latest_published_at":74},"Software","software",93,"2026-09-30T21:41:11.000Z",{"name":76,"slug":77,"count":78,"latest_published_at":79},"Startups","startups",90,"2026-10-01T21:55:22.000Z",{"name":81,"slug":82,"count":83,"latest_published_at":84},"Gaming","gaming",53,"2026-10-02T02:50:39.000Z",{"name":86,"slug":87,"count":88,"latest_published_at":89},"General","general",50,"2026-09-30T21:37:54.000Z",{"name":91,"slug":92,"count":93,"latest_published_at":94},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":96,"slug":97,"count":98,"latest_published_at":99},"How-To","how-to",7,"2026-10-01T09:00:00.000Z"]