[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-researchers-blend-two-ai-training-methods-into-one-sliding-scale":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},8621,"researchers-blend-two-ai-training-methods-into-one-sliding-scale","Researchers Blend Two AI Training Methods Into One Sliding Scale","A new technique called Interpolated Policy Distillation lets AI developers dial between fast, messy training data and slow, clean training data.","A new paper offers a dial for training smaller AI models to imitate larger ones, instead of forcing developers to choose one fixed method.\n\nTraining a compact 'student' model to copy a bigger 'teacher' model is called distillation. Developers can either use teacher-written examples ('off-policy'), which are polished but drift from how the student actually talks, or student-generated attempts corrected by the teacher ('on-policy'), which fit the student better but include real mistakes. Researchers introduce Interpolated Policy Distillation (IPD), which blends the teacher's and student's next-word predictions at every single step, using an adjustable coefficient to set how much weight goes to each. Because checking the teacher's prediction at every step is normally too slow, the team built a speculative-decoding shortcut that gets the same result faster.\n\nThis matters because most teams currently treat off-policy and on-policy distillation as an either-or choice, or bolt them together with a crude two-stage process. IPD gives a single tunable setting instead, and the paper reports it beating both pure approaches, the two-stage combo, and other recent methods that stitch together segments from each source, on text and multimodal reasoning tests.\n\nIt is one arXiv paper with its own benchmarks, not an industry standard. Whether the gains hold up on messier production tasks and larger models remains untested.","[\"ai\",\"model-distillation\",\"llm-training\",\"research\"]","2026-09-30T04:00:00.000Z","2026-09-30T15:27:44.419Z","2026-09-30T15:27:51.019Z","published",null,[],"ai",[24,26,27,28],"model-distillation","llm-training","research",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.37170",0,{"sections":35},[36,39,43,47,52,57,62,67,72,76,81,86,91,96],{"name":37,"slug":24,"count":38,"latest_published_at":18},"AI",5147,{"name":40,"slug":41,"count":42,"latest_published_at":18},"Security","security",788,{"name":44,"slug":45,"count":46,"latest_published_at":18},"Policy","policy",417,{"name":48,"slug":49,"count":50,"latest_published_at":51},"Deals","deals",284,"2026-09-29T21:00:00.000Z",{"name":53,"slug":54,"count":55,"latest_published_at":56},"Hardware","hardware",194,"2026-09-29T13:16:04.000Z",{"name":58,"slug":59,"count":60,"latest_published_at":61},"Science","science",154,"2026-09-28T13:19:18.000Z",{"name":63,"slug":64,"count":65,"latest_published_at":66},"Consumer Tech","consumer-tech",142,"2026-09-29T18:38:03.000Z",{"name":68,"slug":69,"count":70,"latest_published_at":71},"Software","software",91,"2026-09-25T20:55:00.000Z",{"name":73,"slug":74,"count":75,"latest_published_at":18},"Dev Tools","dev-tools",90,{"name":77,"slug":78,"count":79,"latest_published_at":80},"Startups","startups",83,"2026-09-29T21:51:36.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"General","general",49,"2026-09-28T16:44:57.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"Gaming","gaming",48,"2026-09-25T18:35:21.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]