[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-how-to-pick-the-right-layer-for-robot-ai-models":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},8460,"how-to-pick-the-right-layer-for-robot-ai-models","How to Pick the Right Layer for Robot AI Models","A new study finds one mid-network layer choice matters more than fancy multi-layer fusion for teaching vision-language-action models to control robots.","Robots that see, read instructions, and act are usually built by bolting an action module onto a frozen vision-language model. New research says most teams are guessing at the wrong part of that hookup.\n\nResearchers tested three pretrained vision-language backbones on two robot-manipulation benchmarks, LIBERO and CALVIN, checking whether it's better to tap a single layer of the frozen model for the action head or fuse several layers together. Across 54 configurations combining three fusion methods and three layer-selection strategies, 47 did worse than the best single-layer setup. The catch: the best layer differs by backbone and benchmark, so there's no universal answer, and testing every layer by training a full policy is expensive.\n\nThe useful part is a shortcut. The team found that a proxy metric called InfoNCE, run without any robot training, reliably flags which layer will perform well. It cuts the compute needed to pick a layer by 9 to 33 times versus training and testing every option, and it shrinks the performance gap versus the best possible pick from 17.89 percentage points, if you'd just grabbed the deepest layer, down to 3.71 points.\n\nMulti-layer fusion has been a popular default in robotics and multimodal AI on the assumption that more information helps. This paper is a data point against reflexively fusing layers when you could instead spend a little compute finding the one layer that works.","[\"robotics\",\"vision-language-action\",\"machine-learning\",\"arxiv\"]","2026-09-30T04:00:00.000Z","2026-09-30T05:04:06.580Z","2026-09-30T05:04:17.984Z","published",null,[],"ai",[26,27,28,29],"robotics","vision-language-action","machine-learning","arxiv",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.36118",0,{"sections":36},[37,40,44,48,53,58,63,68,73,78,83,88,93,98],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",5028,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",780,{"name":45,"slug":46,"count":47,"latest_published_at":18},"Policy","policy",417,{"name":49,"slug":50,"count":51,"latest_published_at":52},"Deals","deals",284,"2026-09-29T21:00:00.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":57},"Hardware","hardware",194,"2026-09-29T13:16:04.000Z",{"name":59,"slug":60,"count":61,"latest_published_at":62},"Science","science",154,"2026-09-28T13:19:18.000Z",{"name":64,"slug":65,"count":66,"latest_published_at":67},"Consumer Tech","consumer-tech",142,"2026-09-29T18:38:03.000Z",{"name":69,"slug":70,"count":71,"latest_published_at":72},"Software","software",91,"2026-09-25T20:55:00.000Z",{"name":74,"slug":75,"count":76,"latest_published_at":77},"Dev Tools","dev-tools",89,"2026-09-29T17:15:00.000Z",{"name":79,"slug":80,"count":81,"latest_published_at":82},"Startups","startups",83,"2026-09-29T21:51:36.000Z",{"name":84,"slug":85,"count":86,"latest_published_at":87},"General","general",49,"2026-09-28T16:44:57.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"Gaming","gaming",48,"2026-09-25T18:35:21.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]