[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-a-fix-for-self-distillation-that-stops-trashing-reasoning":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},3437,"a-fix-for-self-distillation-that-stops-trashing-reasoning","A Fix for Self-Distillation That Stops Trashing Reasoning","Researchers found that standard on-policy self-distillation breaks long chain-of-thought models, and built a cleaner training signal to fix it.","A new training technique aims to stop a common AI fine-tuning method from quietly destroying the reasoning skills it was meant to improve.\n\nOn-policy self-distillation (OPSD) is a method for making language models better at reasoning by having a teacher model supervise the student on the student's own outputs. The catch, according to new research, is that this process consistently degrades long chain-of-thought models — the kind that reason through problems step by step before answering. The researchers traced the problem to the teacher's supervision signal: it is dominated by shortcuts tied to the reference answer rather than transferable reasoning patterns. The model learns to copy surface features of example solutions, not to think.\n\nThe fix involves two steps. First, the researchers built a reference-only teacher — the same model conditioned on the reference answer but not the question — to isolate the unhelpful part of the signal. Subtracting that out leaves the genuinely useful, question-conditioned correction. They then used pointwise mutual information to convert that residual into a training target the student can actually learn from. Across four long-CoT models and two datasets, the method improved on both the base model and standard OPSD without degrading the models' reflective behavior.\n\nThis matters because long-CoT reasoning is increasingly central to frontier AI performance — it is the mechanism behind models that show their work and catch their own mistakes. A fine-tuning method that quietly erodes that capability while appearing to help is a real problem, not a marginal one.\n\nThe broader lesson: supervision signals in LLM training are messier than they look, and what gets optimized is not always what was intended.","[\"ai\",\"machine-learning\",\"llm\",\"research\"]","2026-07-03T04:00:00.000Z","2026-07-03T06:01:21.858Z","2026-07-03T06:01:24.771Z","published",null,[],"ai",[24,26,27,28],"machine-learning","llm","research",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.02234",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"]