{"id":10404,"date":"2026-07-26T06:52:33","date_gmt":"2026-07-25T22:52:33","guid":{"rendered":"https:\/\/infernews.com\/blog\/a-taxonomy-guided-environment-for-multidomain-visual-reasoning\/"},"modified":"2026-07-26T06:52:33","modified_gmt":"2026-07-25T22:52:33","slug":"a-taxonomy-guided-environment-for-multidomain-visual-reasoning","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/a-taxonomy-guided-environment-for-multidomain-visual-reasoning\/","title":{"rendered":"Trace \u7528\u53ef\u9a57\u8b49\u8cc7\u6599\u91cd\u505a\u8996\u89ba\u63a8\u7406\u8a13\u7df4"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/trace-paper-domain-montage.jpg\" alt=\"trace mark\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f88\u591a\u8996\u89ba\u63a8\u7406\u8cc7\u6599\u96c6\u90fd\u53ea\u4ea4\u4ed8\u5716\u7247\u540c\u7b54\u6848\uff0c\u6a21\u578b\u7b54\u5571\u5497\uff0c\u672a\u5fc5\u4ee3\u8868\u63a8\u7406\u904e\u7a0b\u771f\u4fc2\u7ad9\u5f97\u4f4f\u8173\u3002Trace\u628a\u91cd\u9ede\u653e\u5728<strong>\u53ef\u9a57\u8b49 post-training<\/strong>\uff1a\u5b83\u5c6c\u65bc\u4e00\u500b\u8cc7\u6599\u96c6\u517c\u751f\u6210\u74b0\u5883\uff0c\u91dd\u5c0d\u7684\u554f\u984c\u662f\u600e\u6a23\u7a69\u5b9a\u7522\u751f grounded visual-reasoning \u4efb\u52d9\uff0c\u4e26\u4e14\u8b93\u7b54\u6848\u3001\u6a19\u8a3b\u8207\u9a57\u8b49\u6d41\u7a0b\u4e92\u76f8\u5c0d\u5f97\u4e0a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u63a1\u7528\u4e00\u689d\u5f88\u6e05\u6670\u7684\u751f\u6210\u8def\u7dda\uff1adomain \u2192 scene grammar \u2192 task program\u3002\u73fe\u6709\u505a\u6cd5\u5e38\u898b\u662f\u5148\u6709\u4eba\u624b\u6574\u7406\u984c\u76ee\uff0c\u6216\u8005\u7531\u5716\u50cf\u8207\u6587\u5b57\u9b06\u6563\u914d\u5c0d\uff0c\u518d\u7528\u6700\u7d42\u7b54\u6848\u505a\u76e3\u7763\uff1bTrace\u5247\u7528 deterministic seed \u5148\u5efa\u7acb semantic scene state\uff0c\u518d\u7531 task program \u5f9e\u540c\u4e00\u500b\u72c0\u614b\u63a8\u5c0e typed answer\u3001verifier state\uff0c\u6700\u5f8c\u624d\u6e32\u67d3\u5716\u7247\u8207\u63d0\u793a\u3002\u9019\u7a2e shared-state \u8a2d\u8a08\u7684\u5206\u5225\uff0c\u5728\u65bc\u984c\u76ee\u4e0d\u662f\u300c\u751f\u6210\u5b8c\u518d\u88dc\u6a19\u8a3b\u300d\uff0c\u800c\u662f\u5f9e\u6e90\u982d\u5c31\u628a\u5716\u50cf\u3001\u554f\u984c\u3001\u7b54\u6848\u540c execution trace \u7d81\u5b9a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c0d\u7814\u7a76\u5718\u968a\u4f86\u8aaa\uff0c\u9019\u500b\u53d6\u5411\u5f88\u6709\u5438\u5f15\u529b\uff0c\u56e0\u70ba\u5b83\u540c\u6642\u7167\u9867\u8a13\u7df4\u3001\u6aa2\u67e5\u540c\u91cd\u64ad\u3002\u6bcf\u500b\u4f8b\u5b50\u9664\u4e86 image\u3001prompt\u3001typed answer\uff0c\u9084\u6709 image-space annotation\u3001verifier metadata \u540c execution trace\uff1b\u5c0d\u60f3\u505a RLVR\u3001\u5f8c\u8a13\u7df4\u9a57\u8b49\uff0c\u6216\u8005\u60f3\u5206\u6790\u6a21\u578b\u5230\u5e95\u932f\u5728\u89c0\u5bdf\u3001\u8a08\u7b97\u9084\u662f\u898f\u5247\u7406\u89e3\u7684\u4eba\uff0c\u8cc7\u6599\u5bc6\u5ea6\u6bd4\u4e00\u822c benchmark \u9ad8\u5f97\u591a\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6536\u9304 11 \u500b visual domains\u3001277 \u500b scene grammars\u30011,000 \u500b\u4efb\u52d9<\/li>\n<li>\u5df2\u516c\u958b 66,000 \u500b generated examples\uff0c\u4ea6\u63d0\u4f9b Hugging Face dataset \u8207\u6a21\u578b\u6aa2\u67e5\u9ede<\/li>\n<li>\u9a57\u8b49\u4e0d\u53ea\u770b\u6700\u7d42\u7b54\u6848\uff0c\u9084\u4fdd\u7559 verifier state \u8207 replayable execution trace<\/li>\n<li>\u4ee5 Qwen2.5-VL-3B\u3001Qwen2.5-VL-7B \u505a post-training\uff0c\u5169\u500b\u5c3a\u5ea6\u90fd\u6709\u660e\u986f\u63d0\u5347<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6578\u5b57\u4e0a\uff0c\u5b83\u5728 2,000 \u500b\u672a\u898b\u904e\u3001\u4f46\u7531\u540c\u4e00\u6279 task programs \u751f\u6210\u7684\u65b0\u4f8b\u5b50\u4e0a\uff0c\u5c07 Qwen2.5-VL-3B \u7531 24.45 \u63d0\u5347\u5230 41.05\uff0cQwen2.5-VL-7B \u7531 34.25 \u63d0\u5347\u5230 51.55\u3002\u9019\u4e9b\u7d50\u679c\u9996\u5148\u8aaa\u660e Trace\u5c0d\u540c\u5206\u4f48\u6cdb\u5316\u6709\u5e6b\u52a9\uff1b\u9996\u9801\u4ea6\u63d0\u5230\u7528 64,000 \u500b Trace instances \u8a13\u7df4\u5f8c\uff0c\u5c0d 24 \u500b\u5916\u90e8 benchmarks \u7684 macro-average \u4e5f\u6709\u6539\u5584\uff0c\u4f46\u6458\u8981\u8cc7\u8a0a\u672a\u5217\u5b8c\u6574\u5206\u9805\uff0c\u89e3\u8b80\u6642\u4ecd\u8981\u770b\u539f\u59cb\u5831\u544a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Trace\u6700\u9069\u5408\u88ab\u7406\u89e3\u70ba\u4e00\u500b\u7528\u4f86\u5efa\u69cb\u53ef\u6838\u5c0d\u8996\u89ba\u63a8\u7406\u8a13\u7df4\u8cc7\u6599\u7684\u57fa\u790e\u9805\u76ee\uff0c\u800c\u4e0d\u53ea\u662f\u53e6\u4e00\u500b\u51fa\u984c\u5eab\u3002\u5b83\u7684\u53d6\u6368\u4e5f\u5f88\u660e\u78ba\uff1a\u63db\u4f86\u9ad8\u5ea6\u53ef\u9a57\u8b49\u8207\u53ef\u91cd\u64ad\uff0c\u4ee3\u50f9\u662f\u4efb\u52d9\u5206\u4f48\u7531 scene grammar \u540c task program \u660e\u78ba\u754c\u5b9a\uff0c\u8f03\u9069\u5408\u7814\u7a76\u8a13\u7df4\u65b9\u6cd5\u3001\u8a55\u6e2c\u8a2d\u8a08\u540c\u6a21\u578b\u884c\u70ba\u5206\u6790\uff0c\u672a\u5fc5\u7b49\u540c\u81ea\u7136\u4e16\u754c\u7684\u958b\u653e\u5f0f\u8996\u89ba\u7406\u89e3\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/maveryn.github.io\/trace\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/maveryn\/trace\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2607.19790\" rel=\"noopener noreferrer\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u60f3\u8a13\u7df4\u8996\u89ba\u6a21\u578b\u8b58\u5f97\u771f\u4fc2\u63a8\u7406\uff0c\u800c\u5514\u4fc2\u9760\u984c\u578b\u8a18\u61b6\uff0cTrace\u63d0\u4f9b\u5497\u4e00\u5957\u53ef\u91cd\u64ad\u3001\u53ef\u6838\u5c0d\u7684\u8cc7\u6599\u751f\u6210\u65b9\u6cd5\u3002\u5b83\u4e0d\u53ea\u51fa\u984c\uff0c\u9023\u7b54\u6848\u9ede\u6a23\u5f97\u4f86\u90fd\u4e00\u4f75\u4fdd\u7559\u3002<\/p>\n","protected":false},"author":8,"featured_media":10403,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,147,30,119,127,199],"tags":[],"class_list":["post-10404","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-deepseek","category-image","category-119","category-127","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10404","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=10404"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10404\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10403"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10404"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10404"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10404"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}