[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-a-simpler-way-to-steer-ai-reasoning-without-chain-of-thought":10,"sections":40},{"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":30,"tags":31,"sources":35,"feedback":39,"feedback_at":22,"cost_usd":39,"total_tokens":39},10577,"a-simpler-way-to-steer-ai-reasoning-without-chain-of-thought","A Simpler Way to Steer AI Reasoning Without Chain of Thought","COMPASS steers a few attention heads to boost math accuracy, coming close to chain-of-thought results while using far fewer tokens than full reasoning traces.","A new technique squeezes chain-of-thought-level math reasoning out of language models without the token bill.\n\nResearchers built COMPASS, a method that steers a language model's internal activations toward correct answers without asking it to write out step-by-step reasoning. The team found that a surprisingly simple signal works: whether the model's own quick, direct answer attempts are right or wrong. That signal points to a specific direction inside the model's activations, and while it shows up across most attention heads, only a small subset actually respond to being nudged. COMPASS finds those few heads using a scoring method based on model outputs, then steers just those heads at inference time using statistics gathered per head.\n\nTested across three model families and several math benchmarks, COMPASS lifted GSM8K accuracy by 16 percentage points on average and beat other activation-steering approaches the researchers compared it against. It also came close to, though it didn't fully match, the accuracy of full chain-of-thought prompting, while generating 20 to 70 percent fewer tokens. For anyone paying by the token for AI math help, that is the real pitch: most of the benefit, a fraction of the output.\n\nChain-of-thought prompting works by making a model show its work, which is reliable but verbose; COMPASS is a bet that you can get most of the reasoning benefit by just reaching into the model's head rather than making it type out loud.","[\"ai\",\"llm-reasoning\",\"chain-of-thought\",\"model-interpretability\"]","2026-10-07T04:00:00.000Z","2026-10-08T20:35:33.787Z","2026-10-08T20:35:37.524Z","published",null,[24],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"The source paper says COMPASS 'approaches' chain-of-thought accuracy, not that it 'matched' it — fix the body to say it approaches\u002Fnearly matches CoT performance rather than claiming parity that the source doesn't support.","resolved","ai",[30,32,33,34],"llm-reasoning","chain-of-thought","model-interpretability",[36],{"name":37,"url":38},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.07469",0,{"sections":41},[42,46,51,56,61,66,71,76,81,85,90,95,100,105],{"name":43,"slug":30,"count":44,"latest_published_at":45},"AI",6431,"2026-10-07T18:45:00.000Z",{"name":47,"slug":48,"count":49,"latest_published_at":50},"Security","security",902,"2026-10-07T19:53:42.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Policy","policy",474,"2026-10-07T18:23:21.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":60},"Deals","deals",453,"2026-10-07T23:58:31.000Z",{"name":62,"slug":63,"count":64,"latest_published_at":65},"Hardware","hardware",222,"2026-10-07T21:19:54.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Science","science",186,"2026-10-06T21:20:39.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Consumer Tech","consumer-tech",174,"2026-10-07T17:41:41.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Software","software",113,"2026-10-07T18:10:00.000Z",{"name":82,"slug":83,"count":79,"latest_published_at":84},"Startups","startups","2026-10-07T23:36:57.000Z",{"name":86,"slug":87,"count":88,"latest_published_at":89},"Dev Tools","dev-tools",105,"2026-10-07T16:59:11.000Z",{"name":91,"slug":92,"count":93,"latest_published_at":94},"General","general",61,"2026-10-07T22:00:24.000Z",{"name":96,"slug":97,"count":98,"latest_published_at":99},"Gaming","gaming",56,"2026-10-07T12:00:00.000Z",{"name":101,"slug":102,"count":103,"latest_published_at":104},"Reviews","reviews",33,"2026-10-05T11:57:17.000Z",{"name":106,"slug":107,"count":108,"latest_published_at":109},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]