{"id":9992,"date":"2026-07-08T16:59:00","date_gmt":"2026-07-08T08:59:00","guid":{"rendered":"https:\/\/infernews.com\/blog\/llm-as-a-verifier\/"},"modified":"2026-07-08T16:59:00","modified_gmt":"2026-07-08T08:59:00","slug":"llm-as-a-verifier","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/llm-as-a-verifier\/","title":{"rendered":"LLM-as-a-Verifier \u9ede\u6a23\u91cd\u5beb\u4ee3\u7406\u8a55\u5206"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/logo-1.jpg\" alt=\"LLM-as-a-Verifier\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee3\u7406\u7cfb\u7d71\u6700\u5e38\u898b\u7684\u74f6\u9838\uff0c\u5514\u4fc2\u751f\u6210\u5514\u5230\u7b54\u6848\uff0c\u800c\u4fc2\u4f60\u96e3\u4ee5\u77e5\u9053\u5b83\u6bcf\u4e00\u6b65\u5230\u5e95\u505a\u5f97\u5e7e\u597d\u3002LLM-as-a-Verifier\u5c6c\u65bc<strong>\u958b\u6e90\u6846\u67b6<\/strong>\uff0c\u91dd\u5c0d\u7684\u6b63\u662f\u5462\u500b\u554f\u984c\uff1a\u5b83\u5514\u53ea\u70ba\u6700\u7d42\u7d50\u679c\u6253\u4e00\u6b21\u5206\uff0c\u800c\u4fc2\u7528\u66f4\u7d30\u7dfb\u7684\u65b9\u5f0f\u70ba\u5019\u9078\u7b54\u6848\u3001\u884c\u52d5\u6b65\u9a5f\u540c\u4efb\u52d9\u6e96\u5247\u63d0\u4f9b\u53ef\u91cf\u5316\u56de\u994b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709\u505a\u6cd5\u4e0d\u5c11\u4ecd\u7136\u4f9d\u8cf4\u55ae\u6b21\u5224\u65b7\u3001\u7c97\u7c92\u5ea6\u5206\u6578\uff0c\u6216\u8005\u53ea\u770b\u6700\u7d42\u6210\u6557\uff1b\u4f5c\u8005\u8a8d\u70ba\u5462\u7a2e\u56fa\u5b9a\u7bc4\u5f0f\u6703\u5ffd\u7565\u4e0d\u78ba\u5b9a\u6027\uff0c\u4ea6\u96e3\u4ee5\u652f\u63f4 progress tracking \u540c reinforcement learning\u3002LLM-as-a-Verifier\u6539\u7528\u4e09\u500b\u6838\u5fc3\u8a2d\u8a08\u91cd\u7d44\u9a57\u8b49\u6d41\u7a0b\uff1ascore granularity\u3001repeated evaluation \u540c criteria decomposition\uff0c\u4e26\u4e14\u76f4\u63a5\u5c0d LLM score tokens \u7684\u5b8c\u6574 logprob distribution \u53d6\u671f\u671b\u503c\uff0c\u800c\u5514\u4fc2\u53ea\u53d6\u55ae\u4e00\u8f38\u51fa\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5462\u500b\u53d6\u5411\u4ee4\u5b83\u540c\u4e00\u822c judge-style \u8a55\u5206\u5668\u6709\u660e\u986f\u5206\u91ce\u3002\u5b83\u91cd\u9ede\u5514\u5728\u65bc\u7522\u751f\u4e00\u53e5\u8a55\u8a9e\uff0c\u800c\u4fc2\u7522\u751f\u53ef\u53cd\u8986\u6bd4\u8f03\u3001\u53ef\u5206\u89e3\u3001\u53ef\u7d2f\u7a4d\u7684 fine-grained feedback\uff0c\u6240\u4ee5\u53ef\u4ee5\u7528\u65bc Best-of-N selection\u3001pairwise compare\uff0c\u540c\u57cb\u9010\u6b65\u8ffd\u8e64\u4ee3\u7406\u884c\u70ba\u8b8a\u5316\u3002README \u4ea6\u5c55\u793a\u4e86 Python \u5957\u4ef6 <code>llm-verifier<\/code> \u7684\u57fa\u672c\u7528\u6cd5\uff1b\u5b89\u88dd\u65b9\u5f0f\u6709\u63d0\u4f9b\uff0c\u4f46\u66f4\u5b8c\u6574\u7684\u90e8\u7f72\u7d30\u7bc0\u4e3b\u8981\u653e\u5728\u5b98\u65b9\u6587\u4ef6\uff0c\u800c\u67d0\u4e9b\u6d41\u7a0b\u4ea6\u9700\u8981 <code>VERTEX_API_KEY<\/code> \u6216\u53ef\u56de\u50b3 logprobs \u7684 OpenAI-compatible server\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u652f\u63f4\u4efb\u4f55 modality \u7684\u9a57\u8b49\u6846\u67b6\uff0c\u5b9a\u4f4d\u6bd4\u55ae\u4e00 benchmark judge \u66f4\u5ee3<\/li>\n<li>\u65b9\u6cd5\u6838\u5fc3\u662f\u7d30\u7c92\u5ea6\u8a55\u5206\u3001\u91cd\u8907\u9a57\u8b49\u3001\u6309 criteria \u62c6\u89e3\u6e96\u5247<\/li>\n<li>\u53ef\u76f4\u63a5\u7528\u65bc selection\u3001compare\u3001track\uff0c\u540c\u6642\u9023\u5230 reinforcement learning<\/li>\n<li>\u5b98\u65b9\u5217\u51fa Terminal-Bench V2\u3001SWE-Bench Verified\u3001MedAgentBench\u3001RoboRewardBench \u7b49\u7d50\u679c<\/li>\n<li>\u76f8\u95dc\u6a21\u578b\u8207\u670d\u52d9\u5305\u62ec Qwen\/Qwen3.5-9B\u3001Qwen3-8B\u3001Gemini 2.5 Pro\uff0c\u4ee5\u53ca OpenAI-compatible server<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8868\u73fe\u4e0a\uff0c\u9805\u76ee\u8072\u7a31\u5728\u591a\u500b agentic benchmarks \u9054\u5230 state-of-the-art\uff0c\u5305\u62ec Terminal-Bench V2 86.5%\u3001SWE-Bench Verified 78.2%\u3001RoboRewardBench 87.4%\u3001MedAgentBench 73.3%\uff0c\u4ea6\u63d0\u5230\u5728 LIBERO \u914d\u5408 SAC \u5fae\u8abf <code>pi 0<\/code> policy \u6642\uff0csample efficiency \u7d04\u9ad8 1.8 \u500d\u3002\u5462\u985e\u6578\u5b57\u53cd\u6620\u5b83\u8f03\u9069\u5408\u7814\u7a76\u5718\u968a\u3001\u4ee3\u7406\u5e73\u53f0\u958b\u767c\u8005\uff0c\u540c\u57cb\u60f3\u5c07\u8a55\u4f30\u8a0a\u865f\u63a5\u5165\u8a13\u7df4\u8ff4\u8def\u7684\u4eba\uff1b\u55ae\u7d14\u53ea\u60f3\u8981\u4e00\u500b\u6700\u7d42\u5206\u6578\u7684\u5718\u968a\uff0c\u672a\u5fc5\u9700\u8981\u7528\u5230\u5b83\u6574\u5957\u9a57\u8b49\u5c3a\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/llm-as-a-verifier.com\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/llm-as-a-verifier\/llm-as-a-verifier\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2607.05391\" rel=\"noopener noreferrer\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7576\u4ee3\u7406\u5514\u53ea\u8981\u7b54\u5f97\u5571\uff0c\u4ef2\u8981\u4e00\u6b65\u6b65\u505a\u5f97\u7a69\uff0c\u7c97\u7565\u6253\u5206\u5df2\u7d93\u5514\u5920\u3002LLM-as-a-Verifier\u628a\u8a55\u4f30\u62c6\u7d30\uff0c\u4ee4\u6311\u9078\u3001\u8ffd\u8e64\u540c\u8a13\u7df4\u90fd\u66f4\u6709\u4f9d\u64da\u3002<\/p>\n","protected":false},"author":8,"featured_media":9991,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,140,179,153,116,144,152,191,199],"tags":[],"class_list":["post-9992","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-gemini","category-nvidia","category-openai","category-agentic","category-medical","category-python","category-anthropic","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9992","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=9992"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9992\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9991"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9992"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9992"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9992"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}