[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-study-finds-the-timing-bug-breaking-ai-self-training":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},11358,"study-finds-the-timing-bug-breaking-ai-self-training","Study Finds the Timing Bug Breaking AI Self-Training","Researchers studying how AI models train against earlier versions of themselves found a scheduling fix that prevents sudden, irreversible collapse.","A new study pinpoints exactly when AI self-training goes wrong, and it has nothing to do with how old the teacher is.\n\nIn self on-policy distillation, a student language model trains against a teacher that is really just an earlier snapshot of itself. Researchers ran controlled scheduling experiments on Qwen3-8B to figure out what makes that setup stable. They found that freezing the teacher completely between updates, what they call isolation periods, is the property that matters, not how stale the teacher's weights are. They also identified a failure mode they call state-oblivious collapse: a refresh schedule that looks optimal in short test runs can catastrophically fail over longer training, because a fixed clock can freeze a student's temporary bad stumble into the teacher in one irreversible step. That is a different, sharper problem than the slow contamination seen with exponential-moving-average teacher updates.\n\nThis matters because teams tuning these schedules on quick test runs have no way to see the failure coming. The collapse only shows up once training runs long enough, by which point the damage is baked in. That is a real trap for anyone scaling self-distillation setups without re-testing at production length.\n\nThe researchers' fix, Consolidation-Gated Teacher Refresh, replaces the clock with a checkpoint: only move the teacher once there is evidence the student has actually improved and its output lengths are not drifting into risky territory. Tested on chemistry, biology, physics, and tool-use tasks with one shared configuration and no per-task retuning, it avoided collapse entirely and posted the best scores across the board. It is a governance tweak, not a new architecture, which may be exactly why it held up everywhere it was tried.","[\"ai training\",\"distillation\",\"reinforcement learning\",\"research\"]","2026-10-12T04:00:00.000Z","2026-10-12T10:06:35.094Z","2026-10-12T10:06:41.478Z","published",null,[],"ai",[26,27,28,29],"ai training","distillation","reinforcement learning","research",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.03532",0,{"sections":36},[37,41,46,51,56,61,66,71,76,81,86,91,96,101],{"name":38,"slug":24,"count":39,"latest_published_at":40},"AI",6967,"2026-10-12T15:20:00.000Z",{"name":42,"slug":43,"count":44,"latest_published_at":45},"Security","security",954,"2026-10-12T14:20:37.000Z",{"name":47,"slug":48,"count":49,"latest_published_at":50},"Policy","policy",496,"2026-10-11T20:06:43.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Deals","deals",492,"2026-10-12T14:55:53.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":60},"Hardware","hardware",237,"2026-10-12T13:24:45.000Z",{"name":62,"slug":63,"count":64,"latest_published_at":65},"Science","science",196,"2026-10-11T15:34:40.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Consumer Tech","consumer-tech",192,"2026-10-12T14:07:13.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Startups","startups",121,"2026-10-12T13:22:34.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Software","software",115,"2026-10-11T07:34:45.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Dev Tools","dev-tools",111,"2026-10-12T11:22:24.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",71,"2026-10-12T16:10:40.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Gaming","gaming",59,"2026-10-09T11:43:43.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"Reviews","reviews",35,"2026-10-12T14:00:00.000Z",{"name":102,"slug":103,"count":104,"latest_published_at":105},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]