{"id":9463,"date":"2026-06-21T05:47:39","date_gmt":"2026-06-20T21:47:39","guid":{"rendered":"https:\/\/infernews.com\/blog\/visually-grounded-thinking\/"},"modified":"2026-06-21T05:47:54","modified_gmt":"2026-06-20T21:47:54","slug":"visually-grounded-thinking","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/visually-grounded-thinking\/","title":{"rendered":"visually_grounded_thinking\uff1a\u8b93 VLM \u63a8\u7406\u540c\u6642\u6307\u5411\u5716\u7247\u8b49\u64da"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/illustration.png\" alt=\"Visually grounded thinking modes\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6642\u4e0d\u5c11 Vision-Language Models\uff08VLMs\uff09\u505a\u8996\u89ba\u63a8\u7406\u6642\uff0c\u901a\u5e38\u53ea\u8f38\u51fa\u6587\u5b57\u601d\u8def\uff0c\u8b49\u64da\u5176\u5be6\u4f86\u81ea\u5716\u7247\u54ea\u4e00\u90e8\u5206\uff0c\u6a21\u578b\u672a\u5fc5\u8b1b\u5f97\u6e05\u3002\u4f5c\u8005\u8a8d\u70ba\u9019\u7a2e text-only reasoning trace \u96e3\u9a57\u8b49\u3001\u4ea6\u96e3\u76e3\u7763\uff0c\u6240\u4ee5\u63d0\u51fa visually grounded thinking\uff1a\u5728\u63a8\u7406\u6587\u5b57\u4e2d\u63d2\u5165 <code>&lt;obj&gt;...&lt;\/obj&gt;<\/code>\uff0c\u76f4\u63a5\u6a19\u793a point \u6216 box \u5ea7\u6a19\uff0c\u5c07\u8a9e\u53e5\u540c\u5716\u7247\u5340\u57df\u7d81\u5728\u4e00\u8d77\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u4e0d\u662f\u55ae\u7d14\u52a0\u6a19\u7c64\u7684\u683c\u5f0f\u6539\u52d5\uff0c\u800c\u662f\u4e00\u5957\u8a13\u7df4\u8207\u8a55\u6e2c\u6d41\u7a0b\u3002\u9805\u76ee\u5148\u7528 open-source counting \u8207 spatial reasoning datasets \u5408\u6210\u5e36\u8996\u89ba\u4f9d\u64da\u7684\u601d\u8def\uff0c\u518d\u7528 SAM3-based grounding agent \u7522\u751f mask supervision\uff0c\u4e4b\u5f8c\u5206\u5225\u7528 SFT \u8207 GRPO \u8a13\u7df4\uff1bRL \u968e\u6bb5\u518d\u9760 grounding-aware reward\uff0c\u4ee5 box IoU \u6216 point F1 \u6aa2\u67e5\u6a21\u578b\u6307\u5411\u7684\u7269\u4ef6\u662f\u5426\u5c0d\u61c9\u6b63\u78ba\u8b49\u64da\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u540c\u985e\u505a\u6cd5\u591a\u6578\u95dc\u6ce8\u7b54\u6848\u5571\u5514\u5571\uff0c\u9019\u500b\u9805\u76ee\u9023\u4e2d\u9593\u5f15\u7528\u7684\u5716\u50cf\u8b49\u64da\u90fd\u8a08\u5206\uff0c\u53d6\u5411\u660e\u986f\u8f03\u91cd\u8996\u53ef\u9a57\u8b49\u6027\u3002\u4e0d\u904e\u4ee3\u50f9\u4ea6\u5f88\u6e05\u695a\uff1a\u8cc7\u6599\u88fd\u4f5c\u3001\u7269\u4ef6\u5c0d\u9f4a\u3001reward routing \u90fd\u66f4\u8907\u96dc\uff0c\u8a13\u7df4\u9580\u6abb\u6bd4\u53ea\u505a\u6587\u5b57 reasoning \u9ad8\uff0c\u8f03\u9069\u5408\u5df2\u7d93\u6709 VLM \u8a13\u7df4\u6d41\u7a0b\u7684\u7814\u7a76\u5718\u968a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 counting benchmarks \u8207 spatial reasoning benchmarks\uff0c\u52a0\u4e0a visually grounded thinking \u7684 Gemma3-4B-IT \u666e\u904d\u512a\u65bc\u539f\u7248\u6a21\u578b\u8207 non-grounded thinking baseline\uff1b\u5728\u90e8\u5206\u7a7a\u9593\u63a8\u7406\u4efb\u52d9\uff0c4B \u7248\u672c\u751a\u81f3\u53ef\u8ffd\u4e0a\u6216\u8d85\u904e Gemma3-27B-IT\u3002\u4f5c\u8005\u4ea6\u6307\u51fa point grounding \u8f03\u9069\u5408 counting\uff0c\u800c box grounding \u5728 spatial tasks \u914d\u5408 grounding rewards \u6548\u679c\u66f4\u7a81\u51fa\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u985e\u578b\u4e0a\uff0c\u5b83\u5c6c\u65bc <strong>VLM \u8a13\u7df4\u6846\u67b6\u52a0\u7814\u7a76\u4ee3\u78bc<\/strong>\uff0c\u91cd\u9ede\u662f\u6539\u5584\u8996\u89ba\u63a8\u7406\u904e\u7a0b\u7f3a\u4e4f\u53ef\u6838\u5c0d\u8b49\u64da\u7684\u554f\u984c\u3002<\/li>\n\n\n\n<li>\u5132\u5b58\u5eab\u5df2\u5206\u958b data_synthesis_pipeline\u3001agent\u3001rl_reward\u3001sft_dataset\u3001VLMEvalKit \u8207 scripts\uff0c\u7d50\u69cb\u7b97\u6e05\u6670\uff0c\u7406\u89e3\u6d41\u7a0b\u6703\u6bd4\u76f4\u63a5\u6539\u6a21\u578b\u6b0a\u91cd\u66f4\u91cd\u8981\u3002<\/li>\n\n\n\n<li>\u90e8\u7f72\u601d\u8def\u504f\u7814\u7a76\u7528\u9014\uff0c\u8f03\u53ef\u80fd\u9700\u8981 Docker \u74b0\u5883\u3001SFT\/RL \u8a13\u7df4\u914d\u7f6e\uff0c\u4ee5\u53ca\u81ea\u5099\u7b97\u529b\uff0c\u800c\u4e0d\u662f\u5373\u88dd\u5373\u7528\u7684\u7d42\u7aef\u5de5\u5177\u3002<\/li>\n\n\n\n<li>\u76f8\u95dc\u6a21\u578b\u8207\u5143\u4ef6\u5305\u62ec Gemma3-4B-IT\u3001Gemma3-27B-IT\u3001SAM3\u3001GRPO\u3001SFT\u3001VLMEvalKit\u3002<\/li>\n\n\n\n<li>\u9069\u5408\u95dc\u6ce8\u53ef\u89e3\u91cb\u8996\u89ba\u63a8\u7406\u3001VQA\u3001counting\u3001spatial reasoning\uff0c\u6216\u8005\u60f3\u628a\u4e2d\u9593\u63a8\u7406\u8b8a\u6210\u53ef\u76e3\u7763\u8a0a\u865f\u7684\u5718\u968a\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/Jun-Kai-Zhang\/visually_grounded_thinking\" rel=\"noopener noreferrer\">https:\/\/github.com\/Jun-Kai-Zhang\/visually_grounded_thinking<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2606.16122\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2606.16122<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9805\u76ee\u628a\u6587\u5b57\u63a8\u7406\u9023\u5230\u5716\u7247\u5340\u57df\uff0c\u4ee4 Vision-Language Models \u7684\u4e2d\u9593\u601d\u8def\u66f4\u6613\u9a57\u8b49\uff0c\u4e5f\u66f4\u6613\u8a13\u7df4\u3002<\/p>\n","protected":false},"author":8,"featured_media":9462,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,140,116,30,38,119,150,76,127,149,188,197,199],"tags":[],"class_list":["post-9463","post","type-post","status-publish","format-standard","hentry","category-133","category-gemini","category-agentic","category-image","category-38","category-119","category-150","category-76","category-127","category-149","category-meta","category-framework","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9463","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=9463"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9463\/revisions"}],"predecessor-version":[{"id":9465,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9463\/revisions\/9465"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9462"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9463"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9463"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9463"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}