[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tag-cross-attention":3},{"tag":4,"articles":9},{"id":5,"name":6,"slug":6,"article_count":7,"description_zh":8,"description_en":8},"19ec5cf0-5c97-48b0-8b6a-646937ddaa05","cross-attention",2,null,[10],{"id":11,"slug":12,"title":13,"summary":14,"category":15,"image_url":16,"cover_image":16,"language":17,"created_at":18},"99ae4e8c-3fb2-4d46-8965-458bd16372c8","steerable-vi-vision-representations-zh","可被文字操控的 ViT 特徵","這篇論文把文字提示塞進 ViT 編碼過程中，讓影像特徵能朝指定物件偏移，同時盡量保留通用視覺用途。","research","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775196412799-wufk.png","zh","2026-04-03T06:06:34.443487+00:00"]