{"id":9161,"date":"2026-06-16T05:49:42","date_gmt":"2026-06-15T21:49:42","guid":{"rendered":"https:\/\/infernews.com\/blog\/icml-2026-smaller-models-are-natural-explorers-for-policy-level-diversity-in-grp\/"},"modified":"2026-06-16T05:51:45","modified_gmt":"2026-06-15T21:51:45","slug":"icml-2026-smaller-models-are-natural-explorers-for-policy-level-diversity-in-grp","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/icml-2026-smaller-models-are-natural-explorers-for-policy-level-diversity-in-grp\/","title":{"rendered":"S2L-PO \u7528\u5c0f\u6a21\u578b\u5e36\u52d5\u5927\u6a21\u578b\u63a8\u7406\u8a13\u7df4"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/method-97ef68303d8e.jpg\" alt=\"S2L-PO method overview\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u9805\u76ee\u4f86\u81ea\u8de8\u6821\u5718\u968a\uff0c\u4f5c\u8005\u5305\u62ec Yiming Ren\u3001Yiran Xu\u3001Zicheng Lin \u7b49\u4eba\uff0c\u901a\u8a0a\u4f5c\u8005\u662f Yu Qiao \u8207 Ruihang Chu\uff1b\u6240\u5c6c\u6a5f\u69cb\u5305\u62ec\u6e05\u83ef\u5927\u5b78\u3001\u4e0a\u6d77\u4eba\u5de5\u667a\u6167\u5be6\u9a57\u5ba4\u3001\u9999\u6e2f\u4e2d\u6587\u5927\u5b78\u53ca\u9999\u6e2f\u57ce\u5e02\u5927\u5b78\u3002\u4ee5\u7814\u7a76\u80cc\u666f\u770b\uff0c\u5718\u968a\u660e\u986f\u96c6\u4e2d\u5728\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u3001\u63a8\u7406\u5f37\u5316\u5b78\u7fd2\u8207\u6578\u5b78\u8a55\u6e2c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u662f\u4e00\u500b\u91dd\u5c0d Group Relative Policy Optimization\uff08GRPO\uff09\u8a13\u7df4\u6d41\u7a0b\u7684\u7814\u7a76\u578b\u6846\u67b6\uff0c\u76ee\u7684\u662f\u63d0\u5347\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u5728\u63a8\u7406\u4efb\u52d9\u4e0a\u7684 rollout diversity\u3002\u73fe\u6709\u505a\u6cd5\u591a\u6578\u9760\u63d0\u9ad8 temperature\uff0c\u5f9e token-level randomness \u589e\u52a0\u8b8a\u5316\uff0c\u4f46\u8ad6\u6587\u6307\u51fa\u9019\u7a2e\u56fa\u5b9a\u7bc4\u5f0f\u5bb9\u6613\u5728\u9577\u63a8\u7406\u93c8\u7d2f\u7a4d\u566a\u97f3\uff0c\u4ee4\u8ecc\u8de1\u8b8a\u5f97\u4e0d\u9023\u8cab\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">S2L-PO\uff08Small-to-Large Policy Optimization\uff09\u63db\u4e86\u4e00\u500b\u89d2\u5ea6\uff1a\u4e0d\u7528\u540c\u4e00\u500b\u5927\u6a21\u578b\u4e0d\u505c\u62bd\u6a23\uff0c\u800c\u662f\u627e\u540c\u5bb6\u65cf\u7684\u8f03\u5c0f\u6a21\u578b\u505a explorer\uff0c\u5148\u7522\u751f\u4e00\u90e8\u5206 qualitatively different reasoning trajectories\uff0c\u518d\u8b93\u5927\u6a21\u578b\u7528\u6df7\u5408 rollout \u4f9d\u7167\u6a19\u6e96 GRPO \u8a13\u7df4\u3002\u4e4b\u5f8c\u518d\u7528 progressive annealing\uff0c\u9010\u6b65\u7531 small-model exploration \u904e\u6e21\u5230 fully on-policy learning\uff0c\u907f\u514d\u4e2d\u9014\u88ab\u5c0f\u6a21\u578b\u80fd\u529b\u4e0a\u9650\u62d6\u6162\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ad6\u6587\u63d0\u4f9b\u7684\u7d50\u679c\u9817\u6709\u8aaa\u670d\u529b\u3002\u4ee5 Qwen3-8B learner \u914d 1.7B explorer \u70ba\u4f8b\uff0cAIME24 Pass@1 \u7531 15.0 \u63d0\u5347\u5230 23.8\uff0cAIME25 Pass@1 \u7531 12.1 \u63d0\u5347\u5230 22.5\uff1bQwen3-14B learner \u914d 4B explorer \u4ea6\u6bd4\u57fa\u7dda GRPO \u9ad8\u3002\u4f5c\u8005\u540c\u6642\u8072\u7a31 rollout compute \u9084\u53ef\u964d\u4f4e\uff0c\u9019\u9ede\u5c0d\u8a13\u7df4\u6210\u672c\u654f\u611f\u7684\u5718\u968a\u5c24\u5176\u6709\u5438\u5f15\u529b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u4f60\u60f3\u6e2c\u8a66\u9019\u500b\u9805\u76ee\uff0c\u8f03\u5408\u7406\u7684\u65b9\u5f0f\u4e0d\u662f\u7576\u4f5c\u5373\u88dd\u5373\u7528\u5de5\u5177\uff0c\u800c\u662f\u628a\u5b83\u8996\u70ba\u4e00\u500b\u8a13\u7df4\u7b56\u7565\u53c3\u8003\uff1a\u5148\u770b\u8ad6\u6587\u8207\u516c\u958b\u6a21\u578b\u8a2d\u5b9a\uff0c\u518d\u6bd4\u8f03\u81ea\u5df1\u624b\u4e0a\u7684 GRPO \u6d41\u7a0b\u662f\u5426\u540c\u6a23\u53d7 rollout \u540c\u8cea\u5316\u5f71\u97ff\u3002\u786c\u4ef6\u9580\u6abb\u4e0d\u7b97\u4f4e\uff0c\u8cc7\u6599\u5217\u51fa 8B \u6a21\u578b\u7d04\u9700 20 GB GPU \u8a18\u61b6\u9ad4\u300114B \u6a21\u578b\u7d04\u9700 32 GB\uff0c\u8f03\u9069\u5408\u7814\u7a76\u4eba\u54e1\u3001\u6a21\u578b\u8a13\u7df4\u5de5\u7a0b\u5e2b\uff0c\u6216\u6b63\u5728\u505a\u6578\u5b78\u63a8\u7406\u5fae\u8abf\u7684\u5718\u968a\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6838\u5fc3\u5224\u65b7<\/strong>\uff1a\u9019\u662f\u6a21\u578b\u8a13\u7df4\u6846\u67b6\uff0c\u4e0d\u662f\u4e00\u822c\u804a\u5929\u61c9\u7528\uff0c\u91cd\u9ede\u5728\u6539\u5584 GRPO \u7684\u63a2\u7d22\u54c1\u8cea<\/li>\n\n\n\n<li><strong>\u65b9\u6cd5\u5dee\u7570<\/strong>\uff1a\u7531 token-level randomness \u8f49\u5411 policy-level diversity\uff0c\u6e1b\u5c11\u9577\u93c8\u63a8\u7406\u5931\u771f<\/li>\n\n\n\n<li><strong>\u4e3b\u8981\u5275\u65b0<\/strong>\uff1a\u7528\u8f03\u5c0f\u540c\u5bb6\u65cf\u6a21\u578b\u5145\u7576 natural explorers\uff0c\u518d\u4ee5 progressive annealing \u6536\u56de\u5927\u6a21\u578b\u4e3b\u5c0e\u6b0a<\/li>\n\n\n\n<li><strong>\u5df2\u5217\u76f8\u95dc\u6a21\u578b<\/strong>\uff1aQwen3-1.7B\u3001Qwen3-4B\u3001Qwen3-8B\u3001Qwen3-14B<\/li>\n\n\n\n<li><strong>\u9069\u5408\u5834\u666f<\/strong>\uff1a\u6578\u5b78\u63a8\u7406\u3001\u53ef\u9a57\u8b49\u734e\u52f5\u8a13\u7df4\u3001\u60f3\u63d0\u5347 RLVR \u8207 GRPO \u6536\u6582\u6548\u7387\u7684\u9805\u76ee<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/qishisuren123\/S2L-PO\" rel=\"noopener noreferrer\">https:\/\/github.com\/qishisuren123\/S2L-PO<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2605.30789\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2605.30789<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6e05\u83ef\u8207\u6e2f\u6821\u5718\u968a\u63d0\u51fa S2L-PO\uff0c\u6539\u7528\u5c0f\u6a21\u578b\u63d0\u4f9b\u591a\u6a23\u63a8\u7406\u8def\u5f91\uff0c\u88dc\u8db3 GRPO \u6613\u6536\u6582\u904e\u5feb\u7684\u554f\u984c\u3002<\/p>\n","protected":false},"author":8,"featured_media":9160,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,171,185,163,164,116,38,152,76,127,160,192,196,197,200],"tags":[],"class_list":["post-9161","post","type-post","status-publish","format-standard","hentry","category-133","category-171","category-qwen","category-163","category-164","category-agentic","category-38","category-python","category-76","category-127","category-160","category-192","category-196","category-framework","category-200"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9161","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=9161"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9161\/revisions"}],"predecessor-version":[{"id":9163,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9161\/revisions\/9163"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9160"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9161"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9161"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9161"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}