{"id":8433,"date":"2026-05-21T19:23:24","date_gmt":"2026-05-21T11:23:24","guid":{"rendered":"https:\/\/infernews.com\/blog\/anti-self-distillation-for-reasoning-rl-via-pointwise-mutual-information\/"},"modified":"2026-05-21T19:23:24","modified_gmt":"2026-05-21T11:23:24","slug":"anti-self-distillation-for-reasoning-rl-via-pointwise-mutual-information","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/anti-self-distillation-for-reasoning-rl-via-pointwise-mutual-information\/","title":{"rendered":"AntiSD \u9ede\u6a23\u6539\u5584\u63a8\u7406\u8a13\u7df4\u76f2\u9ede"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/05\/fig1a_overview-b31245b5f162.jpg\" alt=\"fig1a overview\"><\/figure>\n<p>AntiSD \u4fc2\u4e00\u500b\u7814\u7a76\u578b\u5c08\u6848\uff0c\u4e3b\u529b\u8655\u7406\u8a9e\u8a00\u6a21\u578b\u505a\u6578\u5b78\u63a8\u7406\u6642\u5605\u8a13\u7df4\u504f\u5dee\u3002\u4e00\u822c\u505a\u6cd5\u6703\u53eb\u6a21\u578b\u5411\u4e00\u500b\u300c\u5df2\u77e5\u7b54\u6848\u3001\u66f4\u6709\u63d0\u793a\u300d\u5605\u81ea\u5df1\u5b78\u7fd2\uff0c\u4f46\u4f5c\u8005\u6307\u51fa\uff0c\u5462\u7a2e\u5b89\u6392\u5f80\u5f80\u6703\u4ee4\u6a21\u578b\u66f4\u8457\u91cd\u683c\u5f0f\u5316\u3001\u6536\u5c3e\u5f0f\u5605\u5b57\u8a5e\uff0c\u53cd\u800c\u524a\u5f31\u771f\u6b63\u5e6b\u5230\u9010\u6b65\u601d\u8003\u5605\u4e2d\u9593\u63a8\u7406\u7dda\u7d22\u3002<\/p>\n<p>\u5462\u500b\u5c08\u6848\u6700\u7279\u5225\u5605\u5730\u65b9\uff0c\u4fc2\u5c07\u5e38\u898b\u5605\u81ea\u6211\u84b8\u993e\u65b9\u5411\u53cd\u8f49\u3002\u7c21\u55ae\u8b1b\uff0c\u5514\u518d\u4e00\u5473\u903c\u5b78\u751f\u7248\u672c\u8cbc\u8fd1\u300c\u5df2\u7d93\u77e5\u9053\u7b54\u6848\u300d\u5605\u8001\u5e2b\u7248\u672c\uff0c\u800c\u4fc2\u7528\u4e00\u7a2e\u53d7\u63a7\u5236\u5605\u65b9\u5f0f\u4fdd\u7559\u5169\u8005\u5dee\u7570\uff0c\u7b49\u6a21\u578b\u5514\u6703\u904e\u65e9\u653e\u68c4\u63a2\u7d22\u601d\u8def\uff1b\u540c\u6642\u518d\u52a0\u4e0a\u4e00\u500b\u57fa\u65bc\u4e0d\u78ba\u5b9a\u5ea6\u5605\u958b\u95dc\uff0c\u907f\u514d\u8a0a\u865f\u53bb\u5230\u5f8c\u671f\u5931\u63a7\u3002<\/p>\n<p>\u5be6\u969b\u7747\u6cd5\u4e0a\uff0c\u5462\u500b\u65b9\u6cd5\u5514\u4fc2\u7540\u4e00\u822c\u7528\u5bb6\u5373\u88dd\u5373\u73a9\uff0c\u800c\u4fc2\u8f03\u9069\u5408\u5df2\u7d93\u505a\u7dca\u63a8\u7406\u6a21\u578b\u8a13\u7df4\u3001\u60f3\u6bd4\u8f03\u4e0d\u540c\u5f37\u5316\u5b78\u7fd2\u7b56\u7565\u5605\u4eba\u3002\u95b1\u8b80\u8ad6\u6587\u3001\u914d\u5408 GitHub \u5167\u5605\u5be6\u9a57\u8a2d\u5b9a\u540c W&amp;B \u7d50\u679c\u53bb\u91cd\u73fe\uff0c\u6703\u4fc2\u8f03\u5408\u7406\u5605\u4f7f\u7528\u65b9\u5f0f\uff1b\u91cd\u9ede\u4fc2\u89c0\u5bdf\u8a13\u7df4\u6b65\u6578\u3001\u6700\u7d42\u6e96\u78ba\u7387\uff0c\u540c\u57cb\u6a21\u578b\u55ba\u4e2d\u9593\u63a8\u7406\u5b57\u8a5e\u4e0a\u6709\u5187\u88ab\u904e\u5ea6\u58d3\u7e2e\u3002<\/p>\n<ul>\n<li>\u91dd\u5c0d\u6578\u5b78\u63a8\u7406\u8a13\u7df4\u4e2d\u300c\u7b54\u6848\u5571\uff0c\u4f46\u601d\u8def\u8b8a\u8584\u300d\u5605\u554f\u984c<\/li>\n<li>\u6838\u5fc3\u505a\u6cd5\u4fc2\u53cd\u8f49\u81ea\u6211\u84b8\u993e\u8a0a\u865f\uff0c\u800c\u5514\u4fc2\u6cbf\u7528\u6a19\u6e96\u8cbc\u8fd1\u7b56\u7565<\/li>\n<li>\u8ad6\u6587\u63d0\u5230\u4ee5 pointwise mutual information \u89e3\u91cb\u9ede\u89e3\u65b9\u6cd5\u6709\u6548<\/li>\n<li>\u5728\u591a\u500b 4B \u81f3 30B \u6a21\u578b\u4e0a\uff0c\u64da\u5831\u53ef\u7528\u66f4\u5c11\u8a13\u7df4\u6b65\u6578\u8ffd\u5e73\u6216\u8d85\u904e\u57fa\u7dda<\/li>\n<li>\u76f8\u95dc\u6a21\u578b\u5305\u62ec Qwen3-4B\u3001Qwen3-8B\uff0c\u4ee5\u53ca\u5176\u4ed6\u540c\u7d1a 4B \u81f3 30B \u8a9e\u8a00\u6a21\u578b<\/li>\n<\/ul>\n<p>\u4ee5\u5b9a\u4f4d\u4f86\u8b1b\uff0cAntiSD \u6bd4\u8f03\u4f3c\u4e00\u500b\u7540\u7814\u7a76\u54e1\u540c\u6a21\u578b\u5de5\u7a0b\u5718\u968a\u53c3\u8003\u5605\u8a13\u7df4\u914d\u65b9\uff0c\u800c\u5514\u4fc2\u9762\u5411\u7d42\u7aef\u7528\u6236\u5605\u61c9\u7528\u7a0b\u5f0f\u3002\u5c0d\u65bc\u95dc\u6ce8 AIME\u3001HMMT\u3001BeyondAIME \u5462\u985e\u6578\u5b78\u63a8\u7406\u57fa\u6e96\uff0c\u6216\u8005\u6b63\u7528 GRPO \u4e00\u985e\u65b9\u6cd5\u5fae\u8abf\u6a21\u578b\u5605\u5718\u968a\uff0c\u5462\u500b\u5c08\u6848\u63d0\u4f9b\u5497\u4e00\u500b\u503c\u5f97\u8a8d\u771f\u6bd4\u8f03\u5605\u66ff\u4ee3\u65b9\u5411\u3002<\/p>\n<p><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/FloyedShen\/AntiSD\" rel=\"noopener noreferrer\">https:\/\/github.com\/FloyedShen\/AntiSD<\/a><\/p>\n<p><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2605.11609\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2605.11609<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>AntiSD \u91dd\u5c0d\u6578\u5b78\u63a8\u7406\u8a13\u7df4\u5165\u9762\u5e38\u898b\u7684\u300c\u5b78\u5f97\u5feb\u4f46\u8ad7\u5f97\u6dfa\u300d\u554f\u984c\uff0c\u63d0\u51fa\u4e00\u500b\u5e7e\u53cd\u76f4\u89ba\u4f46\u6709\u5be6\u8b49\u652f\u6301\u7684\u65b9\u6cd5\u3002<\/p>\n","protected":false},"author":8,"featured_media":8432,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[133,189,197],"tags":[],"class_list":["post-8433","post","type-post","status-publish","format-standard","hentry","category-133","category-189","category-framework"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8433","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=8433"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8433\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/8432"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=8433"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=8433"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=8433"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}