{"id":9341,"date":"2026-06-18T04:23:40","date_gmt":"2026-06-17T20:23:40","guid":{"rendered":"https:\/\/infernews.com\/blog\/loopcoder\/"},"modified":"2026-06-18T04:24:05","modified_gmt":"2026-06-17T20:24:05","slug":"loopcoder","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/loopcoder\/","title":{"rendered":"LoopCoder\uff1a\u53ea\u591a\u8dd1\u4e00\u8f2a\uff0c\u6210\u6548\u53cd\u800c\u66f4\u597d"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/pasted-86d4ed6893f9.jpg\" alt=\"Only Loop Once: gain\u2013cost trade-off in Parallel Loop Transformers\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">LoopCoder-v2 \u662f\u4e00\u500b\u57fa\u65bc Parallel Loop Transformers\uff08PLT\uff09\u7684\u7a0b\u5f0f\u78bc\u6a21\u578b\u7cfb\u5217\uff0c\u76ee\u6a19\u662f\u89e3\u6c7a\u300c\u63a8\u7406\u6b65\u6578\u6108\u591a\uff0c\u6210\u672c\u8207\u8868\u73fe\u672a\u5fc5\u540c\u6b65\u4e0a\u5347\u300d\u7684\u554f\u984c\u3002\u50b3\u7d71 Looped Transformers \u6703\u900f\u904e\u91cd\u8907\u5171\u4eab\u5340\u584a\u53bb\u589e\u52a0 latent computation\uff0c\u4f46\u6bcf\u591a\u4e00\u8f2a\u90fd\u6703\u62c9\u9ad8\u5ef6\u9072\u548c KV-cache \u8a18\u61b6\u9ad4\uff1bPLT \u5247\u7528 Cross-Loop Position Offsets\uff08CLP\uff09\u548c Shared-KV Gated Sliding-Window Attention\uff08G-SWA\uff09\u628a\u6210\u672c\u58d3\u4f4e\uff0c\u8b93\u8ff4\u5708\u6578\u8b8a\u6210\u53ef\u4ee5\u8abf\u6574\u7684\u8a2d\u8a08\u53c3\u6578\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u9805\u76ee\u76f4\u63a5\u62c6\u89e3\u300c\u591a\u8dd1\u5e7e\u8f2a\u5230\u5e95\u503c\u4e0d\u503c\u5f97\u300d\u3002\u4f5c\u8005\u7528 gain\u2013cost \u89d2\u5ea6\u5206\u6790 loop count\uff1a\u984d\u5916\u4e00\u8f2a\u53ef\u4ee5\u5e36\u4f86\u8868\u793a\u66f4\u65b0\uff0c\u4f46 CLP \u4e5f\u6703\u5f15\u5165\u4f4d\u7f6e\u4e0d\u5339\u914d\u7684\u6210\u672c\uff1b\u5169\u908a\u4e00\u5c0d\u6bd4\uff0c\u5c31\u80fd\u89e3\u91cb\u70ba\u4f55 LoopCoder-v2 \u5728\u5f88\u591a\u60c5\u6cc1\u4e0b\u662f\u5169\u8f2a\u6700\u597d\uff0c\u800c\u4e0d\u662f\u6108\u591a\u6108\u597d\u3002\u9019\u7a2e\u5206\u6790\u65b9\u5f0f\u6bd4\u55ae\u770b\u5206\u6578\u66f4\u6709\u53c3\u8003\u50f9\u503c\uff0c\u56e0\u70ba\u5b83\u628a\u6548\u679c\u5347\u964d\u548c\u5167\u90e8\u6a5f\u5236\u9023\u5728\u4e00\u8d77\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f9e\u7d50\u679c\u770b\uff0cLoopCoder-v2 \u7684 7B \u7248\u672c\u5728\u591a\u500b\u7a0b\u5f0f\u76f8\u95dc\u6e2c\u8a66\u90fd\u6709\u660e\u986f\u6539\u5584\uff0c\u5c24\u5176\u662f SWE-bench Verified \u7531 43.0 \u5347\u5230 64.4\uff0cMulti-SWE \u7531 14.0 \u5347\u5230 31.0\uff0cTerminal-Bench \u4ea6\u6709\u63d0\u5347\u3002\u76f8\u53cd\uff0c\u4e09\u8f2a\u6216\u56db\u8f2a\u6642\u5206\u6578\u660e\u986f\u56de\u843d\uff0c\u8868\u793a\u9019\u500b\u9805\u76ee\u4e0d\u662f\u55ae\u7d14\u9760\u300c\u52a0\u66f4\u591a\u8a08\u7b97\u300d\u63db\u8868\u73fe\uff0c\u800c\u662f\u5b58\u5728\u4e00\u500b\u8f03\u6e05\u6670\u7684\u6700\u4f73\u9ede\u3002\u4f5c\u8005\u4ea6\u7528 hidden-state dynamics\u3001attention evolution \u548c output distribution shift \u53bb\u4f50\u8b49\u7b2c\u4e8c\u8f2a\u5e36\u4f86\u4e3b\u8981\u589e\u76ca\uff0c\u4e4b\u5f8c\u7684\u8f2a\u6b21\u591a\u6578\u53ea\u6703\u589e\u52a0\u5197\u9918\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u4f60\u60f3\u627e\u7684\u662f\u53ef\u76f4\u63a5\u8dd1\u7684\u6a21\u578b\uff0c\u9019\u500b\u9805\u76ee\u63d0\u4f9b\u4e86 Hugging Face \u4e0a\u7684 7B \u6b0a\u91cd\uff0c\u80fd\u900f\u904e Transformers \u8f09\u5165\u5f8c\u505a\u6587\u672c\u751f\u6210\u6216\u7a0b\u5f0f\u78bc\u4efb\u52d9\u6e2c\u8a66\u3002\u9069\u5408\u95dc\u6ce8 code generation\u3001code reasoning\u3001agentic software engineering\u3001tool-use \u7684\u4eba\uff0c\u4e5f\u9069\u5408\u60f3\u7814\u7a76 test-time compute scaling\u3001\u6a21\u578b\u63a8\u7406\u6548\u7387\uff0c\u6216\u60f3\u6bd4\u8f03 loop count \u5c0d\u8868\u73fe\u5f71\u97ff\u7684\u8b80\u8005\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e3b\u8981\u985e\u578b\u662f<strong>\u6a21\u578b\u7814\u7a76\u9805\u76ee<\/strong>\uff0c\u540c\u6642\u5305\u542b\u8a55\u6e2c\u8207\u63a8\u7406\u5206\u6790<\/li>\n\n\n\n<li>\u6838\u5fc3\u7d50\u8ad6\u662f\uff1a<strong>\u5169\u8f2a\u901a\u5e38\u662f\u6700\u4f73\u5e73\u8861\u9ede<\/strong>\uff0c\u4e09\u8f2a\u4ee5\u4e0a\u53ef\u80fd\u53cd\u800c\u62d6\u4f4e\u8868\u73fe<\/li>\n\n\n\n<li>CLP \u4ee4\u5e73\u884c\u8ff4\u5708\u53ef\u884c\uff0cG-SWA \u5247\u628a KV-cache \u6210\u672c\u7dad\u6301\u5728\u8fd1\u4e4e\u56fa\u5b9a\u6c34\u5e73<\/li>\n\n\n\n<li>7B \u7248\u672c\u5728 SWE-bench Verified\u3001Multi-SWE\u3001Terminal-Bench\u3001BFCL \u7b49\u6e2c\u8a66\u90fd\u6709\u8f03\u5b8c\u6574\u7d50\u679c<\/li>\n\n\n\n<li>\u9069\u5408\u7528\u4f86\u5206\u6790\u7a0b\u5f0f\u78bc\u6a21\u578b\u3001\u4ee3\u7406\u5f0f\u4efb\u52d9\uff0c\u4ee5\u53ca\u6e2c\u8a66\u968e\u6bb5\u7b97\u529b\u5206\u914d<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/CSJianYang\/LoopCoder\" rel=\"noopener noreferrer\">https:\/\/github.com\/CSJianYang\/LoopCoder<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2606.18023\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2606.18023<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>LoopCoder-v2 \u7814\u7a76 PLT \u5728\u6e2c\u8a66\u968e\u6bb5\u7684\u8ff4\u5708\u6b21\u6578\u9078\u64c7\uff0c\u986f\u793a\u5169\u8f2a\u5f80\u5f80\u6bd4\u4e09\u8f2a\u4ee5\u4e0a\u66f4\u7a69\u5b9a\u3002\u5c0d\u5beb\u7a0b\u5f0f\u3001\u5de5\u5177\u547c\u53eb\u548c\u4ee3\u7406\u5f0f\u4efb\u52d9\uff0c\u9019\u500b\u9805\u76ee\u63d0\u4f9b\u4e86\u66f4\u5be6\u969b\u7684\u53d6\u6368\u53c3\u8003\u3002<\/p>\n","protected":false},"author":8,"featured_media":9340,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,116,24,38,162,152,109,76,146,197],"tags":[],"class_list":["post-9341","post","type-post","status-publish","format-standard","hentry","category-133","category-agentic","category-aisoftware","category-38","category-ai-productions","category-python","category-rag","category-76","category-146","category-framework"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9341","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=9341"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9341\/revisions"}],"predecessor-version":[{"id":9343,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9341\/revisions\/9343"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9340"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9341"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9341"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9341"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}