{"id":9829,"date":"2026-07-03T19:47:40","date_gmt":"2026-07-03T11:47:40","guid":{"rendered":"https:\/\/infernews.com\/blog\/the-code-and-data-for-quot-perceptionrubrics-calibrating-multimodal-evaluation-t\/"},"modified":"2026-07-03T19:47:40","modified_gmt":"2026-07-03T11:47:40","slug":"the-code-and-data-for-quot-perceptionrubrics-calibrating-multimodal-evaluation-t","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/the-code-and-data-for-quot-perceptionrubrics-calibrating-multimodal-evaluation-t\/","title":{"rendered":"PerceptionRubrics \u9ede\u51fa\u591a\u6a21\u614b\u8a55\u6e2c\u76f2\u9ede"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/pasted-ec914e407867.jpg\" alt=\"Performance Comparison\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">PerceptionRubrics \u662f\u4e00\u500b\u591a\u6a21\u614b\u8a55\u6e2c\u6846\u67b6\u517c\u8cc7\u6599\u96c6\uff0c\u4e3b\u529b\u6aa2\u67e5 Multimodal Large Language Models \u662f\u5426\u771f\u6b63\u770b\u6e05\u5716\u7247\u5167\u5bb9\uff0c\u800c\u5514\u4fc2\u53ea\u4fc2\u5728\u50b3\u7d71 benchmark \u62ff\u5230\u9ad8\u5206\u3002\u5b83\u8981\u89e3\u6c7a\u7684\u554f\u984c\u5f88\u76f4\u63a5\uff1a\u73fe\u6709 caption \u8a55\u6e2c\u5e38\u7528 holistic semantic matching \u6216\u5e73\u5747\u5206\uff0c\u5bb9\u6613\u628a\u56b4\u91cd\u932f\u8aa4\u6c96\u6de1\uff0c\u4f46\u4eba\u985e\u95b1\u8b80\u7d50\u679c\u6642\uff0c\u95dc\u9375\u4e8b\u5be6\u4e00\u932f\uff0c\u6574\u9ad4\u8f38\u51fa\u5df2\u7d93\u672a\u5fc5\u53ef\u4fe1\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f5c\u8005\u628a\u820a\u6709\u7bc4\u5f0f\u62c6\u958b\u91cd\u505a\uff0c\u6539\u7528 atomic auditing\uff0c\u628a\u6bcf\u5f35\u5716\u5206\u89e3\u6210\u53ef\u6838\u5be6\u7684\u7d30\u9805\uff0c\u518d\u5206\u6210 <strong>Must-Right<\/strong> \u8207 <strong>Easy-Wrong<\/strong> \u5169\u689d rubric \u6d41\u3002Must-Right \u91dd\u5c0d\u5fc5\u8981\u4e8b\u5be6\uff0cEasy-Wrong \u91dd\u5c0d\u6a21\u578b\u5e38\u898b\u7684\u7d30\u7bc0\u907a\u6f0f\u3001\u5e7b\u89ba\u6216\u8aa4\u5224\uff1b\u518d\u914d\u5408 gated scoring\uff0c\u53ea\u8981\u5fc5\u8981\u8996\u89ba\u4e8b\u5be6\u51fa\u932f\uff0c\u5c31\u6703\u88ab\u660e\u986f\u6263\u5206\uff0c\u800c\u5514\u4fc2\u88ab\u5176\u4ed6\u5c0f\u5206\u6578\u5e73\u5747\u63a9\u84cb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8cc7\u6599\u898f\u6a21\u65b9\u9762\uff0c\u9805\u76ee\u63d0\u4f9b 1,038 \u5f35 information-dense images\uff0c\u540c\u8d85\u904e 10,000 \u689d instance-specific rubrics\uff0c\u4f86\u6e90\u662f\u7528 Circular Peer-Review \u5efa\u7acb\u7684 Golden Captions\uff0c\u518d\u84b8\u993e\u6210\u8a55\u6e2c\u898f\u5247\u3002\u8986\u84cb\u7bc4\u570d\u5305\u62ec natural scenes\u3001OCR documents\u3001GUIs\u3001charts\u3001STEM\u3001logic puzzles \u540c creative\/cultural images\uff0c\u660e\u986f\u504f\u5411\u9ad8\u8cc7\u8a0a\u5bc6\u5ea6\u3001\u5bb9\u6613\u51fa\u73fe\u611f\u77e5\u5931\u771f\u7684\u5834\u666f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6e2c\u8a66\u65b9\u5f0f\u4e0d\u7b97\u8907\u96dc\uff1a\u9019\u500b GitHub \u5132\u5b58\u5eab\u4e3b\u8981\u63d0\u4f9b evaluation code \u548c data\uff0c\u8f03\u9069\u5408\u7814\u7a76\u5718\u968a\u3001\u6a21\u578b\u958b\u767c\u8005\uff0c\u6216\u8005\u9700\u8981\u6bd4\u8f03\u591a\u500b MLLMs \u8868\u73fe\u7684\u4eba\uff0c\u628a\u6a21\u578b\u8f38\u51fa\u7684 captions \u5c0d\u7167 rubric \u8a08\u5206\u3002\u5b83\u4e0d\u662f\u90e8\u7f72\u7d66\u7d42\u7aef\u7528\u5bb6\u7684\u61c9\u7528\u7a0b\u5f0f\uff0c\u800c\u662f\u62ff\u4f86\u9a57\u8b49\u6a21\u578b\u5728\u5716\u50cf\u7406\u89e3\u4efb\u52d9\u5230\u5e95\u7a69\u4e0d\u7a69\uff1b\u4f7f\u7528\u524d\u4ea6\u8981\u63a5\u53d7\u4e00\u9ede\uff0c\u9019\u985e\u66f4\u56b4\u683c\u7684\u8a55\u5206\u6703\u4ee4\u6a21\u578b\u6210\u7e3e\u6bd4\u50b3\u7d71 leaderboard \u66f4\u96e3\u770b\uff0c\u4f46\u8a3a\u65b7\u50f9\u503c\u66f4\u9ad8\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6838\u5fc3\u53d6\u5411\u662f\u7531 holistic semantic matching \u8f49\u5411 atomic auditing<\/li>\n<li><strong>Must-Right<\/strong> \u8207 <strong>Easy-Wrong<\/strong> \u76f4\u63a5\u5c0d\u61c9\u95dc\u9375\u4e8b\u5be6\u8207\u5e38\u72af\u7d30\u932f<\/li>\n<li><strong>gated scoring<\/strong> \u5f37\u8abf\u300c\u95dc\u9375\u932f\u4e00\u9805\u5c31\u8981\u53cd\u6620\u51fa\u4f86\u300d<\/li>\n<li>\u8cc7\u6599\u96c6\u4e2d\u5728 GUIs\u3001\u6587\u4ef6\u3001\u5716\u8868\u7b49\u9ad8\u5bc6\u5ea6\u8996\u89ba\u4efb\u52d9<\/li>\n<li>\u9069\u5408\u7528\u4f86\u6bd4\u8f03 20+ \u4e3b\u6d41 MLLMs \u7684\u611f\u77e5\u53ef\u9760\u6027\uff0c\u800c\u5514\u53ea\u4fc2\u6bd4\u8f03\u5e73\u5747\u5206<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u9805\u76ee\u6307\u51fa\u6a21\u578b\u7d93\u5e38\u80fd\u8fa8\u8a8d\u96f6\u788e\u5143\u7d20\uff0c\u537b\u672a\u80fd\u540c\u6642\u6eff\u8db3\u591a\u500b\u95dc\u9375\u8996\u89ba\u7d04\u675f\uff0c\u5c24\u5176\u5728 GUIs\u3001documents \u540c structured charts \u66f4\u660e\u986f\u3002README \u8207 supporting context \u4ea6\u63d0\u5230\u66fe\u8a55\u6e2c 20+ \u4e3b\u6d41 MLLMs\uff0c\u5305\u62ec GPT-5.5\uff1b\u4e0d\u904e\u9019\u500b\u5132\u5b58\u5eab\u91cd\u9ede\u4ecd\u7136\u662f\u8a55\u6e2c\u6846\u67b6\u672c\u8eab\uff0c\u800c\u5514\u4fc2\u63a8\u51fa\u65b0\u6a21\u578b\uff0c\u6240\u4ee5\u8f03\u503c\u5f97\u7559\u610f\u7684\u662f\u5b83\u600e\u6a23\u66b4\u9732 perception brittleness\uff0c\u800c\u4e0d\u662f\u55ae\u4e00\u6392\u884c\u699c\u540d\u6b21\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/huggingface.co\/papers\/2606.28322\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/M1chaelPeng\/PerceptionRubrics\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2606.28322\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5462\u500b\u9805\u76ee\u7528\u66f4\u8cbc\u8fd1\u4eba\u985e\u89c0\u611f\u7684\u65b9\u5f0f\uff0c\u91cd\u65b0\u6aa2\u67e5\u591a\u6a21\u614b\u6a21\u578b\u6709\u7121\u770b\u932f\u91cd\u9ede\u3002\u5b83\u5514\u4fc2\u518d\u9b25\u5e73\u5747\u5206\uff0c\u800c\u4fc2\u8ffd\u67e5\u95dc\u9375\u4e8b\u5be6\u6709\u7121\u7b54\u932f\u3002<\/p>\n","protected":false},"author":8,"featured_media":9828,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,176,140,147,153,119,196,199],"tags":[],"class_list":["post-9829","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-176","category-gemini","category-deepseek","category-openai","category-119","category-196","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9829","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=9829"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9829\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9828"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9829"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9829"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9829"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}