[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-tiny-reasoning-model-reveals-answers-before-it-can-use-them":10,"sections":35},{"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":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},5425,"tiny-reasoning-model-reveals-answers-before-it-can-use-them","Tiny Reasoning Model Reveals Answers Before It Can Use Them","A tiny recurrent-depth reasoning model shows internal probes can spot the right answer before its behavior ever demonstrates the skill, researchers find.","A small AI model appears to know the answer before it can actually use it - and that is a problem for how researchers measure learning.\n\nResearchers trained a 30-million-parameter recurrent-depth reasoning model inside a closed, rule-defined toy world, teaching it the same reasoning task two ways: through symbolic notation and through plain verbal descriptions. On the symbolic version, the model mastered three-step (three-hop) reasoning in 70 training epochs. On the verbal version, the identical skill took 13,055 epochs - a 186.5-fold gap. Once three-hop reasoning landed verbally, four-hop reasoning clicked fast, in just 8 more epochs, even though held-out four-hop test performance had been stuck near zero for the entire 13,055-epoch verbal grind. Separately, the team ran linear probes on the model's hidden states and found they could predict the model's eventual answer above chance well before the model could reliably produce that answer through normal behavior, and this pattern held even when they swapped the training surface.\n\nThis matters because most claims about what an AI model \"has learned\" rest on exactly these two signals: benchmark scores and hidden-state readouts. The study shows the two do not agree, and sometimes point in opposite directions - a model can look incompetent on the task while already encoding the right answer internally. The researchers also tried to track that internal signal's growth across training and found the attempt was not clean by design: the probe only counts a case as eligible once the model behaves correctly, so the population being measured keeps changing as the model's behavior changes.\n\nThe paper stops short of claiming the model was quietly \"reasoning\" all along - it argues you cannot tell from behavior or probes alone, and that answering the question needs causal intervention experiments, not just better readouts. Worth remembering next time a benchmark chart is presented as proof a model has learned to think.","[\"ai interpretability\",\"reasoning models\",\"ai research\",\"arxiv\"]","2026-08-18T04:00:00.000Z","2026-08-18T19:40:52.286Z","2026-08-18T19:41:04.163Z","published",null,[],"ai",[26,27,28,29],"ai interpretability","reasoning models","ai research","arxiv",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.16085",0,{"sections":36},[37,41,45,50,55,60,65,70,75,79,84,89,94,99],{"name":38,"slug":24,"count":39,"latest_published_at":40},"AI",3293,"2026-08-20T04:00:00.000Z",{"name":42,"slug":43,"count":44,"latest_published_at":40},"Security","security",435,{"name":46,"slug":47,"count":48,"latest_published_at":49},"Policy","policy",210,"2026-08-19T09:32:27.000Z",{"name":51,"slug":52,"count":53,"latest_published_at":54},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":56,"slug":57,"count":58,"latest_published_at":59},"Hardware","hardware",140,"2026-08-19T18:25:42.000Z",{"name":61,"slug":62,"count":63,"latest_published_at":64},"Consumer Tech","consumer-tech",95,"2026-08-18T16:05:00.000Z",{"name":66,"slug":67,"count":68,"latest_published_at":69},"Science","science",90,"2026-08-19T18:41:02.000Z",{"name":71,"slug":72,"count":73,"latest_published_at":74},"Software","software",73,"2026-08-18T07:51:50.000Z",{"name":76,"slug":77,"count":78,"latest_published_at":18},"Dev Tools","dev-tools",69,{"name":80,"slug":81,"count":82,"latest_published_at":83},"Startups","startups",47,"2026-08-19T19:13:46.000Z",{"name":85,"slug":86,"count":87,"latest_published_at":88},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":90,"slug":91,"count":92,"latest_published_at":93},"General","general",33,"2026-08-18T22:18:13.000Z",{"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"]