{"id":8182,"date":"2026-05-14T07:38:23","date_gmt":"2026-05-13T23:38:23","guid":{"rendered":"https:\/\/infernews.com\/blog\/?p=8182"},"modified":"2026-05-14T07:39:16","modified_gmt":"2026-05-13T23:39:16","slug":"poise","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/poise\/","title":{"rendered":"POISE\uff1a\u8a9e\u8a00\u6a21\u578b\u7528\u81ea\u5df1\u72c0\u614b\u505a\u5f37\u5316\u5b78\u7fd2"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/05\/poise-overview-c2fbd57cc113.jpg\" alt=\"Og image\" \/><\/figure>\n<p><a href=\"https:\/\/holi-lab.github.io\/POISE\/\">POISE<\/a> \u662f\u4e00\u500b\u7528\u65bc\u8a9e\u8a00\u6a21\u578b\u5f37\u5316\u5b78\u7fd2\u7684\u65b9\u6cd5\uff0c\u91cd\u9ede\u662f\u8b93\u6a21\u578b\u5728\u751f\u6210\u7b54\u6848\u6642\uff0c\u76f4\u63a5\u5229\u7528\u81ea\u5df1\u5df2\u6709\u7684\u96b1\u85cf\u72c0\u614b\u8207\u71b5\u7b49\u8a0a\u865f\uff0c\u4f30\u8a08\u56de\u5831\u57fa\u7dda\u3002\u7c21\u55ae\u4f86\u8aaa\uff0c\u5b83\u5e0c\u671b\u6a21\u578b\u300c\u4e00\u908a\u4f5c\u7b54\uff0c\u4e00\u908a\u5224\u65b7\u81ea\u5df1\u8868\u73fe\u5927\u6982\u61c9\u8a72\u6709\u5e7e\u597d\u300d\uff0c\u5f9e\u800c\u66f4\u6709\u6548\u66f4\u65b0\u8a13\u7df4\u65b9\u5411\u3002<\/p>\n<p>\u5be6\u969b\u4f7f\u7528\u4e0a\uff0c\u9019\u985e\u65b9\u6cd5\u7279\u5225\u9069\u5408\u6709\u53ef\u9a57\u8b49\u7d50\u679c\u7684\u4efb\u52d9\uff0c\u4f8b\u5982\u6578\u5b78\u984c\u6700\u7d42\u7b54\u6848\u662f\u5426\u6b63\u78ba\u3002\u50b3\u7d71\u505a\u6cd5\u901a\u5e38\u8981\u984d\u5916\u8a13\u7df4\u4e00\u500b critic \u6a21\u578b\uff0c\u6216\u8005\u5c0d\u540c\u4e00\u63d0\u793a\u505a\u591a\u6b21 rollout \u4f86\u4f30\u57fa\u7dda\uff1bPOISE \u5247\u6539\u70ba\u91cd\u7528 actor \u672c\u8eab\u751f\u6210\u6642\u5df2\u8a08\u7b97\u51fa\u7684\u8cc7\u8a0a\uff0c\u7406\u8ad6\u4e0a\u53ef\u6e1b\u5c11\u8a13\u7df4\u6210\u672c\u3002<\/p>\n<p>\u5b83\u7684\u4e3b\u8981\u5275\u65b0\uff0c\u5728\u65bc\u628a\u6a21\u578b\u5167\u90e8\u8868\u5fb5\u91cd\u65b0\u7d0d\u5165 RL \u66f4\u65b0\u6d41\u7a0b\uff0c\u800c\u4e0d\u662f\u4f9d\u8cf4\u7368\u7acb\u7684\u5927\u578b\u8a55\u5206\u5668\u3002\u7db2\u7ad9\u5167\u5bb9\u6307\u51fa\uff0c\u9019\u500b\u4f30\u503c\u5668\u6703\u96a8\u653f\u7b56\u8b8a\u5316\u6301\u7e8c\u66f4\u65b0\uff0c\u914d\u5408\u7576\u524d\u53ca\u8fd1\u671f rollout\uff0c\u76ee\u6a19\u662f\u4ee4\u7dda\u4e0a\u5b78\u7fd2\u66f4\u7a69\u5b9a\u3002<\/p>\n<ul>\n<li>\u4ee5 actor \u81ea\u8eab\u96b1\u85cf\u72c0\u614b\u505a value estimation<\/li>\n<li>\u4e0d\u9700\u8981\u7368\u7acb critic \u6a21\u578b<\/li>\n<li>\u6e1b\u5c11\u540c\u4e00 prompt \u7684\u984d\u5916\u53d6\u6a23\u958b\u92b7<\/li>\n<li>\u9069\u5408\u6709\u660e\u78ba\u5c0d\u932f\u9a57\u8b49\u7684\u4efb\u52d9<\/li>\n<li>\u8457\u91cd\u66f4\u8f15\u91cf\u8207\u7a69\u5b9a\u7684 RL \u8a13\u7df4\u6d41\u7a0b<\/li>\n<\/ul>\n<p>\u5728\u521d\u6b65\u8a55\u4f30\u65b9\u9762\uff0cPOISE \u5728 held-out \u7684 DAPO-Math rollouts \u4e0a\uff0c\u5167\u90e8\u72c0\u614b\u63a2\u91dd\u7684 Pearson r \u70ba 0.870\u3001MAE \u70ba 0.141\u3002\u5c0d\u6bd4\u4e00\u500b\u7531 Qwen3-4B \u5fae\u8abf\u800c\u6210\u7684 critic\uff0c\u5176 r \u70ba 0.676\u3001MAE \u70ba 0.262\uff0c\u986f\u793a actor \u5167\u90e8\u8a0a\u865f\u81f3\u5c11\u5728\u9019\u9805\u57fa\u6e96\u4e2d\u5177\u6709\u4e0d\u932f\u7684\u9810\u6e2c\u80fd\u529b\u3002<\/p>\n<p>\u53d7\u60e0\u5de5\u4f5c\u5305\u62ec\u6578\u5b78\u89e3\u984c\u3001\u53ef\u7a0b\u5f0f\u9a57\u8b49\u63a8\u7406\uff0c\u4ee5\u53ca\u5176\u4ed6\u80fd\u4ee5\u7d50\u679c\u56de\u994b\u4f5c\u8a13\u7df4\u7684\u8a9e\u8a00\u6a21\u578b\u6d41\u7a0b\u3002\u4e0d\u904e\u76ee\u524d\u9801\u9762\u5c55\u793a\u5167\u5bb9\u4ee5\u65b9\u6cd5\u6982\u5ff5\u8207\u521d\u6b65\u7d50\u679c\u70ba\u4e3b\uff0c\u8f03\u5b8c\u6574\u7684\u6cdb\u5316\u8868\u73fe\u8207\u4e0d\u540c\u4efb\u52d9\u4e0a\u7684\u6700\u7d42\u6548\u76ca\uff0c\u4ecd\u8981\u7559\u610f\u8ad6\u6587\u5f8c\u7e8c\u516c\u958b\u5206\u6790\u3002<\/p>\n\n\n<h2 class=\"wp-block-heading\">\u5be6\u9a57\u7d50\u679c<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th class=\"has-text-align-left\" data-align=\"left\">\u6a21\u578b<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u65b9\u6cd5<\/th><th class=\"has-text-align-left\" data-align=\"left\">Avg@32 (\u6578\u5b78\u63a8\u7406)<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u8a13\u7df4\u6642\u9593 (B200 GPU)<\/th><\/tr><\/thead><tbody><tr><td>Qwen3-4B<\/td><td>DAPO<\/td><td>0.508<\/td><td>49 \u5c0f\u6642<\/td><\/tr><tr><td>Qwen3-4B<\/td><td>POISE<\/td><td><strong>0.500<\/strong><\/td><td><strong>36 \u5c0f\u6642<\/strong>&nbsp;<a rel=\"noreferrer noopener\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2605.07579\"><\/a><\/td><\/tr><tr><td>DeepSeek-R1-Distill-Qwen-1.5B<\/td><td>DAPO<\/td><td>0.296<\/td><td>24 \u5c0f\u6642<\/td><\/tr><tr><td>DeepSeek-R1-Distill-Qwen-1.5B<\/td><td>POISE<\/td><td><strong>0.303<\/strong><\/td><td><strong>18 \u5c0f\u6642<\/strong>&nbsp;<a rel=\"noreferrer noopener\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2605.07579\"><\/a><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>POISE \u5728\u6578\u5b78\u63a8\u7406\u57fa\u6e96 (AMC23\/24, AIME24\/25\/26, HMMT25, BRUMO25) \u4e0a\u9054\u5230\u8207 DAPO \u76f8\u7576\u7684\u6027\u80fd\uff0c\u4f46\u8a08\u7b97\u6210\u672c\u66f4\u4f4e \u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>POISE \u8b93\u8a9e\u8a00\u6a21\u578b\u76f4\u63a5\u5229\u7528\u751f\u6210\u6642\u7684\u5167\u90e8\u8a0a\u865f\u4f30\u503c\uff0c\u6e1b\u5c11\u984d\u5916\u8a55\u8ad6\u5668\u8207\u91cd\u8907\u53d6\u6a23\u958b\u92b7\u3002\u5c0d\u6578\u5b78\u7b54\u6848\u9a57\u8b49\u9019\u985e\u4efb\u52d9\u5c24\u5176\u503c\u5f97\u7559\u610f\u3002<\/p>\n","protected":false},"author":8,"featured_media":8181,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[150,127,160],"tags":[],"class_list":["post-8182","post","type-post","status-publish","format-standard","hentry","category-150","category-127","category-160"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8182","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=8182"}],"version-history":[{"count":2,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8182\/revisions"}],"predecessor-version":[{"id":8186,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8182\/revisions\/8186"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/8181"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=8182"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=8182"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=8182"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}