[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-vision-language-models-dont-know-what-they-cant-see":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},10707,"vision-language-models-dont-know-what-they-cant-see","Vision Language Models Don't Know What They Can't See","A new benchmark finds leading vision-language models trust misleading camera angles over physical reality, and rarely ask for a better view.","Vision-language models can describe a room. They are much worse at admitting when the room they're describing is the wrong one, or hidden behind a wall.\n\nA new benchmark called SPATIALUNCERTAIN tests eight open- and closed-source vision-language models on two kinds of unreliable observation: evidence blocked by occlusion, and evidence distorted by a misleading camera angle. The researchers also checked whether models could tell when they needed a different viewpoint to answer confidently. Across the board, models did not grow more cautious as occlusion removed more of the scene. Under a conflicting perspective, their answers increasingly matched the flat, projected appearance of objects rather than the actual, unchanged 3D layout. A look inside the models found a likely cause: the 2D projection is easy to extract from their internal representations, while the real 3D relationship is barely decodable at all.\n\nThat asymmetry is the real finding. It means these models aren't failing at geometry so much as failing to notice that geometry and appearance have split apart. Spatial-reasoning benchmarks have mostly graded whether a model gets the answer right, not whether it knows when its view can't support an answer at all - a gap that matters the moment a model is doing navigation, inspection, or any task where the camera angle is someone else's choice.\n\nPrompting and fine-tuning did not fix the problem. The one thing that helped was obvious in hindsight: giving the model an actually better viewpoint worked far better than handing it depth data about the same bad one. Even then, models that could pick a better viewpoint when asked to still often failed to recognize they needed one in the first place. In other words, these systems can be more decisive than they are self-aware - a trait that reads fine in a chatbot and less fine in anything that has to move through a room it can't fully see.","[\"ai\",\"computer-vision\",\"spatial-reasoning\",\"benchmarks\"]","2026-10-07T04:00:00.000Z","2026-10-09T05:29:05.704Z","2026-10-09T05:29:09.201Z","published",null,[],"ai",[24,26,27,28],"computer-vision","spatial-reasoning","benchmarks",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.30557",0,{"sections":35},[36,40,45,50,55,60,64,69,74,78,83,88,93,98],{"name":37,"slug":24,"count":38,"latest_published_at":39},"AI",6517,"2026-10-08T04:00:00.000Z",{"name":41,"slug":42,"count":43,"latest_published_at":44},"Security","security",911,"2026-10-07T19:53:42.000Z",{"name":46,"slug":47,"count":48,"latest_published_at":49},"Policy","policy",474,"2026-10-07T18:23:21.000Z",{"name":51,"slug":52,"count":53,"latest_published_at":54},"Deals","deals",458,"2026-10-08T10:03:39.000Z",{"name":56,"slug":57,"count":58,"latest_published_at":59},"Hardware","hardware",223,"2026-10-08T08:30:04.000Z",{"name":61,"slug":62,"count":63,"latest_published_at":18},"Science","science",187,{"name":65,"slug":66,"count":67,"latest_published_at":68},"Consumer Tech","consumer-tech",175,"2026-10-08T09:16:05.000Z",{"name":70,"slug":71,"count":72,"latest_published_at":73},"Software","software",113,"2026-10-07T18:10:00.000Z",{"name":75,"slug":76,"count":72,"latest_published_at":77},"Startups","startups","2026-10-07T23:36:57.000Z",{"name":79,"slug":80,"count":81,"latest_published_at":82},"Dev Tools","dev-tools",105,"2026-10-07T16:59:11.000Z",{"name":84,"slug":85,"count":86,"latest_published_at":87},"General","general",61,"2026-10-07T22:00:24.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"Gaming","gaming",56,"2026-10-07T12:00:00.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",33,"2026-10-05T11:57:17.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]