[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-method-corrects-ai-vision-models-spatial-mistakes":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},7567,"new-method-corrects-ai-vision-models-spatial-mistakes","New Method Corrects AI Vision Models' Spatial Mistakes","RelCheck pairs scene-graph analysis with raw geometry to catch when vision-language models misplace objects in a scene, and the numbers back it up.","A new correction tool aims to stop AI vision models from mixing up where things actually are in a photo.\n\nResearchers built RelCheck, a training-free system that reviews and rewrites a vision-language model's output after the fact, rather than retraining the model itself. It targets relational hallucinations specifically - errors about how objects relate to each other in space, as opposed to whether an object exists or what color it is. The system checks two independent sources of evidence: scene-graph relationships predicted by a model called RelTR, and hard geometric rules calculated from each object's bounding box. Those combine with an existing object-verification layer, styled on the Woodpecker correction method, and a language model uses all three layers to rewrite the flawed description.\n\nMost hallucination-correction research has chased object and attribute errors - is there really a dog in the photo, is the shirt really red - while spatial relationships have stayed a blind spot. Tested on LLaVA v1 13B, RelCheck lifted the MME hallucination benchmark score from 585.0 to 630.0, with the sharpest jump on position questions, where accuracy rose from 0.367 to 0.600. An ablation test found both evidence layers contributed independently rather than one doing all the work, a useful sanity check for a method built by stacking components.\n\nIt's a narrow fix for a narrow failure mode, not a cure for AI vision's habit of seeing things that aren't there - and it hasn't been peer reviewed yet, so it still has to hold up outside the lab.","[\"ai\",\"computer vision\",\"hallucinations\",\"research\"]","2026-09-24T04:00:00.000Z","2026-09-24T07:04:42.185Z","2026-09-24T07:04:48.356Z","published",null,[],"ai",[24,26,27,28],"computer vision","hallucinations","research",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.27890",0,{"sections":35},[36,39,43,48,53,58,63,68,73,78,83,88,93,98],{"name":37,"slug":24,"count":38,"latest_published_at":18},"AI",4424,{"name":40,"slug":41,"count":42,"latest_published_at":18},"Security","security",724,{"name":44,"slug":45,"count":46,"latest_published_at":47},"Policy","policy",380,"2026-09-23T22:53:43.000Z",{"name":49,"slug":50,"count":51,"latest_published_at":52},"Deals","deals",227,"2026-09-24T11:08:33.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":57},"Hardware","hardware",174,"2026-09-24T10:10:29.000Z",{"name":59,"slug":60,"count":61,"latest_published_at":62},"Science","science",136,"2026-09-24T09:00:00.000Z",{"name":64,"slug":65,"count":66,"latest_published_at":67},"Consumer Tech","consumer-tech",116,"2026-09-24T00:51:49.000Z",{"name":69,"slug":70,"count":71,"latest_published_at":72},"Software","software",85,"2026-09-23T20:00:00.000Z",{"name":74,"slug":75,"count":76,"latest_published_at":77},"Dev Tools","dev-tools",79,"2026-09-22T22:21:13.000Z",{"name":79,"slug":80,"count":81,"latest_published_at":82},"Startups","startups",66,"2026-09-23T17:28:38.000Z",{"name":84,"slug":85,"count":86,"latest_published_at":87},"Gaming","gaming",45,"2026-09-22T15:35:06.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"General","general",43,"2026-09-21T23:48:56.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",27,"2026-09-22T13:00:00.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]