{"id":13713,"date":"2026-10-21T21:57:33","date_gmt":"2026-10-21T13:57:33","guid":{"rendered":"https:\/\/infernews.com\/blog\/coevo-rsi\/"},"modified":"2026-10-21T22:01:03","modified_gmt":"2026-10-21T14:01:03","slug":"coevo-rsi","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/coevo-rsi\/","title":{"rendered":"CoEvo-RSI\uff1a\u8b93\u6a21\u578b\u8ddf\u4e0a\u6703\u9032\u5316\u7684 Agent Harness"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/CoEvo-RSI-Agent-Harness.jpg\" alt=\"CoEvoX framework\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u7576 Agent \u7684\u57f7\u884c\u6d41\u7a0b\u4e0d\u65b7\u6539\u8b8a\uff0c\u6a21\u578b\u53ef\u80fd\u672a\u5b78\u61c2\u65b0\u8981\u6c42\uff0c\u4e5f\u53ef\u80fd\u56e0 Harness \u9650\u5236\u800c\u7528\u4e0d\u51fa\u539f\u6709\u80fd\u529b\u3002CoEvoX \u5c6c\u65bc LLM agent \u7684\u6a21\u578b\u2014Harness \u5171\u6f14\u6846\u67b6\uff0c\u900f\u904e\u5171\u540c\u6f14\u5316\u6b77\u53f2\u6a94\u6848\uff0c\u5354\u8abf\u6a21\u578b\u6aa2\u67e5\u9ede\u3001\u57f7\u884c\u6846\u67b6\u8207\u5be6\u9a57\u7d50\u679c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u8f2a\u6d41\u7a0b\u7531 meta-agent \u7de8\u8f2f\u65b0 Harness\uff0c\u5f9e\u6b77\u53f2\u6a94\u6848\u9078\u64c7\u6216\u5408\u4f75\u6a21\u578b checkpoint\uff0c\u518d\u7528\u62d2\u7d55\u53d6\u6a23\u8ecc\u8de1\u9032\u884c SFT\uff1b\u53ea\u6709\u901a\u904e\u9a57\u8b49\u7684\u6539\u5584\u7d44\u5408\u624d\u6703\u6649\u5347\u3002TIMR\uff08Turn-level Interface Mismatch Rate\uff09\u7528\u4f86\u8a3a\u65b7\u9010\u56de\u5408\u7684\u4ecb\u9762\u932f\u914d\uff0c\u4ea6\u4ee4\u6a21\u578b\u9078\u64c7\u4e0d\u518d\u53ea\u770b\u55ae\u4e00\u6a21\u578b\u80fd\u529b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u2022 Qwen3.5-9B \u7d93 CoEvoX \u8abf\u6574\u5f8c\uff0c\u90e8\u5206\u5931\u6548\u6b78\u56e0\u6e2c\u8a66\u8d85\u8d8a\u66f4\u5927\u578b\u6a21\u578b<br \/>\n\u2022 \u8a55\u4f30\u6db5\u84cb\u591a\u4ee3\u7406\u5931\u6548\u6b78\u56e0\u3001\u89d2\u8272\u611f\u77e5 text-to-SQL \u53ca\u667a\u80fd\u5bb6\u5c45\u6307\u4ee4<br \/>\n\u2022 Held-out \u6e2c\u8a66\u907f\u514d\u540c\u6642\u53c3\u8207\u8a13\u7df4\u8207\u6f14\u5316\u9078\u64c7<br \/>\n\u2022 \u7d50\u679c\u4e26\u975e\u5168\u9762\u9818\u5148\uff0c\u90e8\u5206\u6f14\u7b97\u6cd5\u751f\u6210\u6216 When \u6307\u6a19\u4ecd\u843d\u5f8c\u5176\u4ed6\u65b9\u6cd5<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5931\u6548\u6b78\u56e0\u7d50\u679c\u4e2d\uff0cCoEvoX \u5728 Hand-Crafted \u7684 Who\u3001Pro \u7684 Who\uff0c\u4ee5\u53ca TraceElephant \u7684 Who \u5206\u5225\u53d6\u5f97 68.39%\u300169.43% \u548c 70.78%\uff0c\u4f46\u5728 TraceElephant When \u53ea\u6709 28.92%\uff0c\u4f4e\u65bc GSPO \u7684 38.36%\u3002\u9019\u53cd\u6620\u5171\u6f14\u5316\u80fd\u6539\u5584\u7279\u5b9a\u5de5\u4f5c\u5206\u5de5\uff0c\u537b\u672a\u5fc5\u5c0d\u6240\u6709\u63a8\u7406\u6b65\u9a5f\u540c\u6a23\u6709\u6548\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9805\u76ee\u8f03\u9069\u5408\u7814\u7a76 recursive self-improving agents\u3001\u6a21\u578b\u8a13\u7df4\u53ca Harness \u8a2d\u8a08\u7684\u5718\u968a\uff0c\u7528\u4f86\u6e2c\u8a66\u6a21\u578b\u8207\u5de5\u4f5c\u6d41\u7a0b\u662f\u5426\u4e92\u76f8\u914d\u5408\u3002\u63d0\u4f9b\u7684\u8cc7\u6599\u6c92\u6709\u5b8c\u6574\u5b89\u88dd\u3001\u57f7\u884c\u6307\u4ee4\u6216\u90e8\u7f72\u6d41\u7a0b\uff0c\u4ea6\u4e0d\u61c9\u76f4\u63a5\u5047\u8a2d\u53ef\u96a8\u610f\u4e0b\u8f09\uff1b\u76ee\u524d\u8f03\u5408\u7406\u7684\u7406\u89e3\u662f\u628a\u5b83\u8996\u70ba\u9644\u6709\u5be6\u9a57\u6846\u67b6\u8207\u8a55\u4f30\u65b9\u5411\u7684\u7814\u7a76\u9805\u76ee\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/SuJQ0722\/CoEvo-RSI\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/huggingface.co\/collections\/nullptr04\/coevo-rsi\" rel=\"noopener noreferrer\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>CoEvoX \u8b93\u6a21\u578b\u8207\u57f7\u884c\u6846\u67b6\u540c\u6b65\u6f14\u5316\uff0c\u6539\u5584 AI Agent \u7684\u80fd\u529b\u8207\u5de5\u4f5c\u6d41\u7a0b\u4e92\u76f8\u932f\u914d\u53ca\u81ea\u6211\u9032\u5316(RSI &#8211; Recursive Self-Improving)\u7684\u554f\u984c\u3002<\/p>\n","protected":false},"author":8,"featured_media":13712,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,116,76,127,185,197],"tags":[],"class_list":["post-13713","post","type-post","status-publish","format-standard","hentry","category-133","category-agentic","category-76","category-127","category-qwen","category-framework"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13713","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=13713"}],"version-history":[{"count":2,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13713\/revisions"}],"predecessor-version":[{"id":13716,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13713\/revisions\/13716"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/13712"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=13713"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=13713"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=13713"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}