
{"id":10406,"date":"2026-07-26T06:56:33","date_gmt":"2026-07-25T22:56:33","guid":{"rendered":"https:\/\/infernews.com\/blog\/riemannian-isometric-policy-optimization-icml-2026\/"},"modified":"2026-07-26T06:56:33","modified_gmt":"2026-07-25T22:56:33","slug":"riemannian-isometric-policy-optimization-icml-2026","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/riemannian-isometric-policy-optimization-icml-2026\/","title":{"rendered":"RIPO \u76f4\u6307 LLM \u5f37\u5316\u5b78\u7fd2\u63a2\u7d22\u5d29\u584c"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/pasted-3f229786226b.jpg\" alt=\"Repository image for Aiolus-X\/RIPO\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8a13\u7df4 LLM \u505a\u9577\u93c8\u63a8\u7406\u6642\uff0c\u6700\u9ebb\u7169\u7684\u4e0d\u53ea\u662f\u7b97\u529b\uff0c\u800c\u662f\u7b56\u7565\u5f88\u5bb9\u6613\u6108\u5b78\u6108\u4fdd\u5b88\uff0c\u6700\u5f8c\u5361\u5728\u5c11\u6578\u9ad8\u6a5f\u7387\u7b54\u6848\u9644\u8fd1\u3002RIPO \u5c6c\u65bc\u4e00\u500b\u9762\u5411 LLM \u5f37\u5316\u5b78\u7fd2\u7684\u6f14\u7b97\u6cd5\u7814\u7a76\u9805\u76ee\uff0c\u91dd\u5c0d\u7684\u6b63\u662f PPO-Clip \u5728\u5f8c\u8a13\u7df4\u968e\u6bb5\u5e38\u898b\u7684 exploration collapse\uff0c\u60f3\u89e3\u6c7a\u7f55\u898b\u4f46\u95dc\u9375\u52d5\u4f5c\u6108\u4f86\u6108\u96e3\u88ab\u63a2\u7d22\u5230\u7684\u554f\u984c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f5c\u8005\u6c92\u6709\u6cbf\u7528\u300c\u518d\u88dc\u5e7e\u500b heuristic\u300d\u7684\u8def\u7dda\uff0c\u800c\u662f\u76f4\u63a5\u6307\u51fa\u820a\u7bc4\u5f0f\u7684\u6838\u5fc3\u932f\u4f4d\uff1aPPO-Clip \u4ee5 Euclidean metric \u91cf\u5ea6 policy discrepancy\uff0c\u4f46 policy \u672c\u8eab\u66f4\u8cbc\u8fd1\u4e00\u500b Riemannian manifold\u3002\u5462\u500b\u5e7e\u4f55\u4e0d\u4e00\u81f4\u6703\u4ee4\u4f4e\u6a5f\u7387\u5340\u57df\u66f4\u65b0\u904e\u4efd\u4fdd\u5b88\u3001\u9ad8\u6a5f\u7387\u5340\u57df\u53c8\u904e\u4efd\u9032\u53d6\uff0c\u6700\u5f8c\u4ee4\u63a2\u7d22\u80fd\u529b\u6536\u7e2e\uff1bRiemannian Isometric Policy Optimization\uff08RIPO\uff09\u5247\u6539\u70ba\u8ffd\u6c42\u7b49\u8ddd\u7684 policy update\uff0c\u5617\u8a66\u540c\u6642\u5b88\u4f4f exploration \u8207 exploitation \u7684\u5e73\u8861\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ad6\u6587\u63cf\u8ff0\u4e2d\uff0cRIPO \u53e6\u4e00\u500b\u91cd\u9ede\u662f bias-variance trade-off \u8f03\u7406\u60f3\uff0c\u4ee4\u512a\u5316\u904e\u7a0b\u66f4\u7a69\u5b9a\u3002\u6210\u6548\u65b9\u9762\uff0c\u5b83\u5728\u4e03\u500b competition-level benchmarks \u4e0a\u90fd\u512a\u65bc\u65e2\u6709 LLM RL algorithms\uff0c\u7576\u4e2d\u5c0d GRPO \u5728 AIME24 \u7684\u63d0\u5347\u6700\u9ad8\u53ef\u9054 60%\uff1b\u9019\u985e\u7d50\u679c\u76f8\u7576\u5438\u5f15\uff0c\u4f46\u4ecd\u7136\u8981\u7559\u610f benchmark \u8207\u8a13\u7df4\u8a2d\u7f6e\u662f\u5426\u80fd\u5b8c\u6574\u8f49\u79fb\u5230\u4f60\u624b\u4e0a\u7684\u6a21\u578b\u8207\u8cc7\u6599\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6838\u5fc3\u6279\u8a55\u5f88\u660e\u78ba\uff1aPPO-Clip \u7684\u5e7e\u4f55\u5047\u8a2d\u4e0d\u9069\u5408 policy update<\/li>\n<li>\u65b9\u6cd5\u91cd\u5fc3\u4e0d\u662f\u52a0\u734e\u52f5\u6280\u5de7\uff0c\u800c\u662f\u91cd\u5beb\u7b56\u7565\u66f4\u65b0\u7684\u5ea6\u91cf\u65b9\u5f0f<\/li>\n<li>\u5c0d\u6578\u5b78\u63a8\u7406\u3001\u9577\u6642\u5e8f\u6c7a\u7b56\u9019\u985e\u8981\u9760\u63a2\u7d22\u627e\u5230\u89e3\u6cd5\u7684\u8a13\u7df4\u5834\u666f\u8f03\u6709\u50f9\u503c<\/li>\n<li>\u6210\u7e3e\u4eae\u773c\uff0c\u4f46\u66f4\u9069\u5408\u6709\u80fd\u529b\u91cd\u8dd1 benchmark \u8207\u8a13\u7df4\u6d41\u7a0b\u7684\u7814\u7a76\u5718\u968a\u9a57\u8b49<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f9e\u63d0\u4f9b\u7684 GitHub \u8cc7\u8a0a\u770b\uff0c\u5132\u5b58\u5eab\u63cf\u8ff0\u6df7\u5165\u4e86 verl \u9019\u500b RL training library \u7684\u5167\u5bb9\uff0c\u56e0\u6b64\u95b1\u8b80\u8207\u90e8\u7f72\u524d\u8981\u5148\u5206\u6e05\uff1aRIPO \u662f\u6f14\u7b97\u6cd5\u8207\u8ad6\u6587\u65b9\u5411\uff0cverl \u5247\u8f03\u50cf\u627f\u8f09 LLM RL \u8a13\u7df4\u6d41\u7a0b\u7684\u958b\u6e90\u57fa\u5efa\u3002\u8f03\u5408\u7406\u7684\u7406\u89e3\u65b9\u5f0f\uff0c\u662f\u628a RIPO \u8996\u70ba\u53ef\u6574\u5408\u9032\u73fe\u6709 LLM RL framework \u7684\u65b0\u7b56\u7565\u66f4\u65b0\u65b9\u6cd5\uff1b\u771f\u6b63\u843d\u5730\u901a\u5e38\u8981\u914d\u5408\u65e2\u6709\u8a13\u7df4\u5eab\u3001GPU \u53e2\u96c6\u914d\u7f6e\uff0c\u4ee5\u53ca\u50cf GRPO\u3001PPO \u4e00\u985e\u5f8c\u8a13\u7df4 dataflow \u4e00\u8d77\u6e2c\u8a66\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/Aiolus-X\/RIPO\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2607.10169\" rel=\"noopener noreferrer\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>RIPO \u4e0d\u662f\u518d\u8abf\u4e00\u500b PPO-Clip \u8b8a\u9ad4\uff0c\u800c\u662f\u76f4\u63a5\u8cea\u7591\u91cf\u5ea6\u7b56\u7565\u5dee\u7570\u7684\u65b9\u6cd5\u3002\u5c0d\u60f3\u63d0\u5347 LLM \u63a8\u7406\u8a13\u7df4\u7a69\u5b9a\u5ea6\u7684\u4eba\uff0c\u9019\u500b\u65b9\u5411\u503c\u5f97\u7559\u610f\u3002<\/p>\n","protected":false},"author":8,"featured_media":10405,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,176,153,150,127,191,196,197,199],"tags":[],"class_list":["post-10406","post","type-post","status-publish","format-standard","hentry","category-133","category-176","category-openai","category-150","category-127","category-anthropic","category-196","category-framework","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10406","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=10406"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10406\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10405"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10406"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10406"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10406"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}