
{"id":13480,"date":"2026-10-16T23:38:28","date_gmt":"2026-10-16T15:38:28","guid":{"rendered":"https:\/\/infernews.com\/blog\/terminal-rsi-recursive-self-improving-terminal-agents\/"},"modified":"2026-10-16T23:38:28","modified_gmt":"2026-10-16T15:38:28","slug":"terminal-rsi-recursive-self-improving-terminal-agents","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/terminal-rsi-recursive-self-improving-terminal-agents\/","title":{"rendered":"Terminal-RSI \u8b93\u7d42\u7aef\u6a5f\u4ee3\u7406\u5faa\u74b0\u6539\u5beb\u7b56\u7565"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/Terminal-RSI.jpg\" alt=\"Og image\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u7d42\u7aef\u6a5f\u4ee3\u7406\u9762\u5c0d\u8907\u96dc\u6307\u4ee4\u6642\uff0c\u5f80\u5f80\u4e0d\u662f\u5b8c\u5168\u4e0d\u61c2\uff0c\u800c\u662f\u5728\u591a\u8f2a\u64cd\u4f5c\u4e2d\u9010\u6b65\u504f\u96e2\u6b63\u78ba\u8def\u5f91\u3002Terminal-RSI \u662f\u4e00\u9805\u7814\u7a76\u9805\u76ee\uff0c\u5229\u7528\u9a57\u8b49\u6210\u529f\u7684 terminal trajectories\uff0c\u8b93 27B \u53c3\u6578\u6a21\u578b\u53cd\u8986\u4fee\u6b63\u81ea\u8eab\u884c\u70ba\uff0c\u6e1b\u5c11\u540c\u985e\u4efb\u52d9\u91cd\u72af\u932f\u8aa4\u7684\u60c5\u6cc1\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9805\u76ee\u628a Recursive Self-Rewrite \u8207 Policy-Refine \u5206\u6210\u5169\u689d\u4e92\u76f8\u914d\u5408\u7684\u6d41\u7a0b\u3002\u6a21\u578b\u5148\u900f\u904e\u4e0d\u540c harness \u7522\u751f\u64cd\u4f5c\u8ecc\u8de1\uff0c\u518d\u7531 terminal verifier \u7be9\u8d70\u5931\u6557\u7d50\u679c\uff1b\u6210\u529f\u8ecc\u8de1\u6703\u4ea4\u7d66 actor \u9032\u884c SFT\uff0ccritic \u5247\u4fdd\u7559\u4e26\u91cd\u65b0\u6821\u6e96\uff0c\u4e4b\u5f8c\u518d\u4ee5 actor\u2013critic PPO \u4e00\u540c\u512a\u5316\u653f\u7b56\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u63a1\u7528 TB-GAE\uff08Turn-Based Generalized Advantage Estimation\uff09\u8ffd\u8e64\u751f\u6210 token \u7684\u8cac\u4efb\uff0c\u89c0\u5bdf\u5167\u5bb9\u7559\u5728\u4e0a\u4e0b\u6587\u4e2d\uff0c\u4fe1\u7528\u5206\u914d\u5247\u6cbf\u7740\u6bcf\u8f2a\u64cd\u4f5c\u5411\u5f8c\u50b3\u905e\u3002\u5b8c\u6210\u4e00\u8f2a\u5f8c\uff0c\u4fdd\u7559\u7684 checkpoint \u6703\u6210\u70ba\u4e0b\u4e00\u8f2a\u905e\u6b78\u6539\u9032\u7684\u8d77\u9ede\uff0c\u800c\u975e\u6bcf\u6b21\u90fd\u7531\u96f6\u958b\u59cb\u8a13\u7df4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6e2c\u8a66\u7d50\u679c\u986f\u793a\uff0cTerminal-Bench 4.0 \u7684\u5206\u6578\u7531 0.00% \u5347\u81f3 19.84%\uff0cTerminal-Bench 2.1 \u5247\u7531 51.70% \u5347\u81f3 86.52%\uff0c\u5206\u5225\u63d0\u5347 19.84 \u500b\u767e\u5206\u9ede\u53ca 34.82 \u500b\u767e\u5206\u9ede\u3002\u5169\u500b\u57fa\u6e96\u7248\u672c\u7684\u4efb\u52d9\u8207\u96e3\u5ea6\u4e0d\u540c\uff0c\u6578\u5b57\u8f03\u9069\u5408\u7528\u4f86\u89c0\u5bdf\u5404\u81ea\u7684\u6539\u9032\u5e45\u5ea6\uff0c\u4e0d\u80fd\u76f4\u63a5\u7576\u4f5c\u540c\u4e00\u689d\u6392\u884c\u699c\u6bd4\u8f03\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u53ea\u4fdd\u7559\u901a\u904e verifier \u7684\u6210\u529f\u8ecc\u8de1<\/li>\n<li>\u4ee5 Recursive Self-Rewrite \u6539\u5beb\u53ef\u91cd\u7528\u7684\u64cd\u4f5c\u7d93\u9a57<\/li>\n<li>\u4ee5 Policy-Refine \u6301\u7e8c\u8abf\u6574 actor\u2013critic \u653f\u7b56<\/li>\n<li>Terminal-RSI-27B \u662f\u76ee\u524d\u5c55\u793a\u7684\u6700\u7d42 checkpoint<\/li>\n<li>Code \u8207 model \u4ecd\u6a19\u793a\u70ba\u5373\u5c07\u63a8\u51fa<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c0d\u9700\u8981\u6a21\u578b\u57f7\u884c\u6a94\u6848\u8655\u7406\u3001\u7a0b\u5f0f\u7de8\u8f2f\u6216\u591a\u6b65\u9a5f shell \u4efb\u52d9\u7684\u5de5\u4f5c\u6d41\uff0c\u9019\u7a2e\u8a13\u7df4\u65b9\u5f0f\u63d0\u4f9b\u4e86\u53e6\u4e00\u689d\u63d0\u5347\u53ef\u9760\u6027\u7684\u65b9\u5411\u3002\u4e0d\u904e\u9805\u76ee\u4ecd\u8655\u65bc\u7814\u7a76\u5c55\u793a\u968e\u6bb5\uff0c\u4ee3\u78bc\u53ca\u6a21\u578b\u5c1a\u672a\u516c\u958b\uff0c\u7d50\u679c\u4ea6\u4e3b\u8981\u96c6\u4e2d\u65bc\u6307\u5b9a\u7684 Terminal-Bench \u8a55\u6e2c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/jyyang26.github.io\/terminal-rsi\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Terminal-RSI \u900f\u904e\u6210\u529f\u8ecc\u8de1\u53cd\u8986\u6539\u5beb\u8207\u5fae\u8abf\u7b56\u7565\uff0c\u63d0\u5347\u6a21\u578b\u5b8c\u6210\u7d42\u7aef\u6a5f\u4efb\u52d9\u7684\u80fd\u529b\u3002<\/p>\n","protected":false},"author":8,"featured_media":13479,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,172,116,76],"tags":[],"class_list":["post-13480","post","type-post","status-publish","format-standard","hentry","category-133","category-172","category-agentic","category-76"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13480","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=13480"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13480\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/13479"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=13480"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=13480"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=13480"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}