{"id":3888,"date":"2025-06-06T11:40:34","date_gmt":"2025-06-06T03:40:34","guid":{"rendered":"https:\/\/www.ihub1.com\/?p=3888"},"modified":"2025-06-06T11:41:51","modified_gmt":"2025-06-06T03:41:51","slug":"qwen%e6%b8%85%e5%8d%8e%e5%9b%a2%e9%98%9f%e9%a2%a0%e8%a6%86%e5%b8%b8%e8%af%86%ef%bc%9a%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%e4%bb%85%e7%94%a820%e5%85%b3%e9%94%aetoken","status":"publish","type":"post","link":"https:\/\/www.ihub1.com\/index.php\/2025\/06\/06\/qwen%e6%b8%85%e5%8d%8e%e5%9b%a2%e9%98%9f%e9%a2%a0%e8%a6%86%e5%b8%b8%e8%af%86%ef%bc%9a%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%e4%bb%85%e7%94%a820%e5%85%b3%e9%94%aetoken\/","title":{"rendered":"Qwen&amp;\u6e05\u534e\u56e2\u961f\u98a0\u8986\u5e38\u8bc6\uff1a\u5927\u6a21\u578b\u5f3a\u5316\u5b66\u4e60\u4ec5\u752820%\u5173\u952etoken\uff0c\u6bd4\u7528\u5168\u90e8token\u8bad\u7ec3\u8fd8\u597d"},"content":{"rendered":"\t\t<div data-elementor-type=\"wp-post\" data-elementor-id=\"3888\" class=\"elementor elementor-3888\">\n\t\t\t\t<div class=\"elementor-element elementor-element-927c05d e-flex e-con-boxed e-con e-parent\" data-id=\"927c05d\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-44da14c elementor-widget elementor-widget-text-editor\" data-id=\"44da14c\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t<p class=\"art_p\"><span>\u8fd1\u671farxiv\u6700\u70ed\u95e8\u8bba\u6587\uff0c<\/span><span>wen&amp;\u6e05\u534eLeapLab<\/span><span>\u56e2\u961f\u6700\u65b0\u6210\u679c\uff1a<\/span><\/p><p class=\"art_p\"><span>\u5728\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u5927\u6a21\u578b\u63a8\u7406\u80fd\u529b\u65f6\uff0c<\/span><span>\u4ec5\u4ec520%\u7684\u9ad8\u71b5token\u5c31\u80fd\u6491\u8d77\u6574\u4e2a\u8bad\u7ec3\u6548\u679c<\/span><span>\uff0c\u751a\u81f3\u6bd4\u7528\u5168\u90e8token\u8bad\u7ec3\u8fd8\u8981\u597d\u3002<\/span><\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"0\" data-puid=\"0\"><img decoding=\"async\" id=\"0\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/517\/w1080h237\/20250605\/2363-aef0b2f9233779d163cdf5192ac5e651.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u56e2\u961f\u7528\u8fd9\u4e2a\u53d1\u73b0\u5728Qwen3-32B\u4e0a\u521b\u9020\u4e86\u65b0\u7684SOTA\u8bb0\u5f55\uff1aAIME\u201924\u4e0a\u8fbe\u523063.5\u5206\uff0cAIME\u201925\u4e0a\u8fbe\u523056.7\u5206\uff0c<\/p><p class=\"art_p\">\u8fd9\u662f600B\u53c2\u6570\u4ee5\u4e0b\u76f4\u63a5\u4ecebase\u6a21\u578b\u8bad\u7ec3\u7684\u6700\u9ad8\u5206\u3002<\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"1\" data-puid=\"0\"><img decoding=\"async\" id=\"1\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/721\/w1080h441\/20250605\/38c2-cce534ff9c04cd892f835f33743f2e80.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u6700\u5927\u54cd\u5e94\u957f\u5ea6\u4ece20k\u5ef6\u957f\u523029k\uff0cAIME\u201924\u7684\u5206\u6570\u66f4\u662f\u98d9\u5347\u5230\u4e8668.1\u5206\u3002<\/p><p class=\"art_p\">\u7ecf\u5178\u7684\u4e8c\u516b\u6cd5\u5219\uff08\u6216\u5e15\u7d2f\u6258\u6cd5\u5219\uff09\u6307\u51fa\uff0c\u901a\u5e3880%\u7684\u7ed3\u679c\u753120%\u7684\u5173\u952e\u56e0\u7d20\u9a71\u52a8\uff0c\u4f46\u5269\u4e0b80%\u4e5f\u662f\u4e0d\u80fd\u8f7b\u6613\u820d\u5f03\u7684\u3002<\/p><p class=\"art_p\">\u4f46\u662f\u5728\u5927\u6a21\u578b\u5f3a\u5316\u5b66\u4e60\u8fd9\u91cc\uff0c80%\u4f4e\u71b5token\u4e0d\u4ec5\u53ef\u4ee5\u820d\u5f03\uff0c\u751a\u81f3\u8fd8\u53ef\u80fd\u8d77\u526f\u4f5c\u7528\uff0c\u6240\u4ee5\u8fd9\u7bc7\u8bba\u6587\u88ab\u547d\u540d\u4e3a\u201c\u8d85\u8d8a\u4e8c\u516b\u6cd5\u5219\u201d\u3002<\/p><p class=\"art_p\"><span>\u6b64<\/span><span>\u5916\uff0c\u56e2\u961f\u8fd8\u4ecetoken\u71b5\u7684\u89d2\u5ea6\u63a2\u7a76\u4e86RL\u5bf9LLM\u7684\u4e3b\u8981\u5f71\u54cd\uff0c\u5e76\u8fdb\u4e00\u6b65\u8ba8\u8bba\u4e86RL\u4e0eSFT\u7684\u533a\u522b\u3001LLM RL\u7684\u7279\u6b8a\u6027\u4e0eclip-higher\u76f8\u8f83\u4e8eentropy bonus\u7684\u4f18\u52bf\u3002<\/span><\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"2\" data-puid=\"0\"><img decoding=\"async\" id=\"2\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/629\/w1080h349\/20250605\/9d31-74d9102239e79176b1f6c3979bd360bd.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u63ed\u5f00Chain-of-Thought\u7684\u71b5\u5206\u5e03\u5bc6\u7801<\/p><p class=\"art_p\">\u8981\u7406\u89e3\u8fd9\u9879\u7814\u7a76\uff0c\u9700\u8981\u5148\u4ece\u4e00\u4e2a\u6709\u8da3\u7684\u89c2\u5bdf\u8bf4\u8d77\uff1a<\/p><p class=\"art_p\"><span>\u56e2\u961f\u53d1\u73b0\uff0c\u5f53\u5927\u6a21\u578b\u8fdb\u884c\u94fe\u5f0f\u601d\u8003\uff08Chain-of-Thought\uff09\u63a8\u7406\u65f6\uff0ctoken\u7684\u71b5\u5206\u5e03\u5448\u73b0\u51fa\u4e00\u4e2a\u72ec\u7279\u7684\u6a21\u5f0f\uff1a<\/span><span>\u5927\u90e8\u5206token\u7684\u71b5\u90fd\u5f88\u4f4e\uff0c\u53ea\u6709\u5c11\u6570token\u8868\u73b0\u51fa\u9ad8\u71b5\u7279\u5f81<\/span><span>\u3002<\/span><\/p><p class=\"art_p\">\u5177\u4f53\u6765\u8bf4\uff0c\u8d85\u8fc750%\u7684token\u71b5\u503c\u4f4e\u4e8e0.01\uff0c\u800c\u53ea\u670920%\u7684token\u71b5\u503c\u5927\u4e8e0.672\u3002<\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"3\" data-puid=\"0\"><img decoding=\"async\" id=\"3\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/471\/w1080h191\/20250605\/0fb1-8de5f9163805cbffd3bb05f5f70bf68a.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u66f4\u6709\u610f\u601d\u7684\u662f\uff0c\u8fd9\u4e9b\u9ad8\u71b5token\u548c\u4f4e\u71b5token\u5728\u529f\u80fd\u4e0a\u6709\u7740\u672c\u8d28\u533a\u522b\u3002\u9ad8\u71b5token\u901a\u5e38\u626e\u6f14\u7740\u201c\u903b\u8f91\u8fde\u63a5\u5668\u201d\u7684\u89d2\u8272\uff0c\u6bd4\u5982\u201cwait\u201d\u3001\u201chowever\u201d\u3001\u201cthus\u201d\u8fd9\u4e9b\u8bcd\uff0c\u5b83\u4eec\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u8d77\u5230\u8f6c\u6298\u3001\u9012\u8fdb\u6216\u56e0\u679c\u8fde\u63a5\u7684\u4f5c\u7528\u3002<\/p><p class=\"art_p\">\u5728\u6570\u5b66\u63a8\u5bfc\u4e2d\uff0c\u201csuppose\u201d\u3001\u201cassume\u201d\u3001\u201cgiven\u201d\u8fd9\u4e9b\u5f15\u5165\u5047\u8bbe\u6216\u6761\u4ef6\u7684\u8bcd\u4e5f\u5c5e\u4e8e\u9ad8\u71b5token\u3002\u800c\u4f4e\u71b5token\u5219\u5f80\u5f80\u662f\u8bcd\u7f00\u3001\u4ee3\u7801\u7247\u6bb5\u6216\u6570\u5b66\u8868\u8fbe\u5f0f\u7684\u7ec4\u6210\u90e8\u5206\uff0c\u8fd9\u4e9b\u5185\u5bb9\u5177\u6709\u5f88\u9ad8\u7684\u786e\u5b9a\u6027\u3002<\/p><p class=\"art_p\"><span>\u56e2\u961f\u5f62\u8c61\u5730\u628a\u8fd9\u4e9b\u9ad8\u71b5token\u79f0\u4e3a<\/span><span>\u5206\u53c9token<\/span><span>\uff08forking tokens\uff09\uff0c<\/span><span>\u5982\u679c\u5c06\u601d\u7ef4\u94fe\u6bd4\u505a\u8d70\u5c71\u8def\uff0c\u9ad8\u71b5token\u5c31\u50cf\u5206\u53c9\u8def\u53e3\uff08fork\uff09\uff0c\u51b3\u5b9a\u4e86\u63a5\u4e0b\u6765\u7684\u65b9\u5411\uff1b\u4f4e\u71b5token\u5c31\u50cf\u4e00\u9762\u662f\u5c71\u4e00\u9762\u662f\u60ac\u5d16\u7684\u5c0f\u8def\uff0c\u53ea\u80fd\u6cbf\u7740\u65e2\u5b9a\u65b9\u5411\u8d70\u4e0b\u53bb\u3002<\/span><\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"4\" data-puid=\"0\"><img decoding=\"async\" id=\"4\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/713\/w1080h433\/20250605\/de22-260fb7b47de8315ed3985c2c2a6e0eba.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u4e3a\u4e86\u9a8c\u8bc1\u8fd9\u4e9b\u5206\u53c9token\u7684\u91cd\u8981\u6027\uff0c\u56e2\u961f\u8fd8\u8bbe\u8ba1\u4e86\u8fd9\u6837\u7684\u5b9e\u9a8c\uff1a\u7ed9\u4e0d\u540c\u7684token\u8bbe\u7f6e\u4e0d\u540c\u7684\u89e3\u7801\u6e29\u5ea6\u3002<\/p><p class=\"art_p\">\u7ed3\u679c\u53d1\u73b0\uff0c\u9002\u5ea6\u63d0\u9ad8\u9ad8\u71b5token\u7684\u6e29\u5ea6\u80fd\u6539\u5584\u63a8\u7406\u6027\u80fd\uff0c\u800c\u964d\u4f4e\u5b83\u4eec\u7684\u6e29\u5ea6\u5219\u4f1a\u5bfc\u81f4\u6027\u80fd\u4e0b\u964d\u3002\u8fd9\u8fdb\u4e00\u6b65\u8bc1\u5b9e\u4e86\u4fdd\u6301\u8fd9\u4e9b\u5173\u952e\u4f4d\u7f6e\u7684\u9ad8\u71b5\u5bf9\u63a8\u7406\u7684\u91cd\u8981\u6027\u3002<\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"5\" data-puid=\"0\"><img decoding=\"async\" id=\"5\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/26\/w886h740\/20250605\/40be-fde34ca049b04403d99279a98ef5cbd2.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u53ea\u752820%\u7684token\uff0c\u6548\u679c\u53cd\u800c\u66f4\u597d<\/p><p class=\"art_p\">\u65e2\u7136\u53d1\u73b0\u4e86\u8fd9\u4e9b\u5206\u53c9token\u7684\u7279\u6b8a\u6027\uff0c\u90a3\u4e48\u5982\u679c\u5728\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u65f6\u53ea\u5173\u6ce8\u8fd9\u4e9b\u5173\u952e\u7684\u5c11\u6570token\u4f1a\u600e\u6837\uff1f<\/p><p class=\"art_p\"><span>\u5728RLVR\uff08Reinforcement Learning with Verifiable Rewards\uff09\u8bad\u7ec3\u4e2d\uff0c<\/span><span>\u53ea\u4fdd\u7559top 20%\u9ad8\u71b5token\u7684\u7b56\u7565\u68af\u5ea6\uff0c\u628a\u5269\u4e0b80%\u7684\u68af\u5ea6\u5168\u90e8\u5c4f\u853d\u6389<\/span><span>\u3002<\/span><\/p><p class=\"art_p\">\u7ed3\u679cQwen3-32B\u4e0a\uff0c\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u4ec5\u6ca1\u6709\u964d\u4f4e\u6027\u80fd\uff0c\u53cd\u800c\u5e26\u6765\u4e86\u663e\u8457\u63d0\u5347\uff1aAIME\u201924\u5206\u6570\u63d0\u53477.71\u5206\uff0cAIME\u201925\u5206\u6570\u63d0\u534711.04\u5206\uff0c\u5e73\u5747\u54cd\u5e94\u957f\u5ea6\u589e\u52a0\u7ea61378\u4e2atoken\u3002<\/p><p class=\"art_p\">\u5728Qwen3-14B\u4e0a\u4e5f\u6709\u7c7b\u4f3c\u7684\u63d0\u5347\u6548\u679c\uff0c\u800c\u5728\u8f83\u5c0f\u7684Qwen3-8B\u4e0a\uff0c\u6027\u80fd\u4e5f\u4fdd\u6301\u7a33\u5b9a\u3002<\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"6\" data-puid=\"0\"><img decoding=\"async\" id=\"6\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/397\/w1080h917\/20250605\/76d9-6830fde28dc65af9f2b07ac5d91a7b9d.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u66f4\u6709\u8da3\u7684\u662f\u53cd\u5411\u5b9e\u9a8c\u7684\u7ed3\u679c\uff1a\u5982\u679c\u53ea\u752880%\u7684\u4f4e\u71b5token\u8bad\u7ec3\uff0c\u6a21\u578b\u6027\u80fd\u6025\u5267\u4e0b\u964d\u3002<\/p><p class=\"art_p\">\u8fd9\u8bf4\u660e\u4f4e\u71b5token\u5bf9\u63a8\u7406\u80fd\u529b\u7684\u63d0\u5347\u8d21\u732e\u5fae\u4e4e\u5176\u5fae\uff0c\u751a\u81f3\u53ef\u80fd\u8d77\u5230\u8d1f\u9762\u4f5c\u7528\u3002<\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"7\" data-puid=\"0\"><img decoding=\"async\" id=\"7\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/742\/w1080h462\/20250605\/7613-e7877a9c4a71265225331b521dc40b5d.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u56e2\u961f\u5206\u6790\u8ba4\u4e3a\uff0c\u8fd9\u53ef\u80fd\u4e0e\u8def\u5f84\u63a2\u7d22\u6709\u5173\u3002\u9ad8\u71b5token\u7531\u4e8e\u5176\u4e0d\u786e\u5b9a\u6027\uff0c\u80fd\u591f\u5e2e\u52a9\u6a21\u578b\u63a2\u7d22\u4e0d\u540c\u7684\u63a8\u7406\u8def\u5f84\u3002\u800c\u4f4e\u71b5token\u8fc7\u4e8e\u786e\u5b9a\uff0c\u53cd\u800c\u9650\u5236\u4e86\u6a21\u578b\u7684\u63a2\u7d22\u80fd\u529b\u3002<\/p><p class=\"art_p\">\u4ece\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u6574\u4f53\u71b5\u53d8\u5316\u4e5f\u80fd\u770b\u51fa\u7aef\u502a\uff1a\u4f7f\u752820%\u9ad8\u71b5token\u8bad\u7ec3\u65f6\uff0c\u6a21\u578b\u4fdd\u6301\u4e86\u8f83\u9ad8\u7684\u6574\u4f53\u71b5\u6c34\u5e73\uff0c\u800c\u4f7f\u7528\u5168\u90e8token\u6216\u53ea\u7528\u4f4e\u71b5token\u65f6\uff0c\u6574\u4f53\u71b5\u6c34\u5e73\u90fd\u4f1a\u4e0b\u964d\u3002<\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"8\" data-puid=\"0\"><img decoding=\"async\" id=\"8\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/411\/w1080h931\/20250605\/5eed-9ce441c2dbea8aee29d645fb89bed07f.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u53e6\u4e00\u4e2a\u91cd\u8981\u53d1\u73b0\u662f\u8fd9\u79cd\u65b9\u6cd5\u5b58\u5728\u660e\u663e\u7684\u89c4\u6a21\u6548\u5e94\uff1a\u6a21\u578b\u8d8a\u5927\uff0c\u53ea\u7528\u9ad8\u71b5token\u8bad\u7ec3\u7684\u4f18\u52bf\u5c31\u8d8a\u660e\u663e\u3002\u572832B\u6a21\u578b\u4e0a\u7684\u63d0\u5347\u6700\u5927\uff0c14B\u6b21\u4e4b\uff0c8B\u6a21\u578b\u63d0\u5347\u6700\u5c0f\u3002\u8fd9\u53ef\u80fd\u662f\u56e0\u4e3a\u66f4\u5927\u7684\u6a21\u578b\u6709\u66f4\u5f3a\u7684\u80fd\u529b\u5229\u7528\u589e\u5f3a\u7684\u63a2\u7d22\u6027\u3002<\/p><p class=\"art_p\">\u66f4\u4ee4\u4eba\u60ca\u559c\u7684\u662f\uff0c\u8fd9\u79cd\u65b9\u6cd5\u8bad\u7ec3\u51fa\u7684\u6a21\u578b\u5728\u57df\u5916\u4efb\u52a1\u4e0a\u4e5f\u8868\u73b0\u4f18\u5f02\u3002\u56e2\u961f\u5728\u6570\u5b66\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\uff0c\u5728\u7f16\u7a0b\u4efb\u52a1LiveCodeBench\u4e0a\u6d4b\u8bd5\u65f6\uff0c\u4ec5\u752820%\u9ad8\u71b5token\u8bad\u7ec3\u7684\u6a21\u578b\u4f9d\u7136\u5927\u5e45\u8d85\u8d8a\u4e86\u4f7f\u7528\u5168\u90e8token\u8bad\u7ec3\u7684\u7248\u672c\u3002<\/p><p class=\"art_p\">\u8fd9\u6697\u793a\u7740\u9ad8\u71b5token\u53ef\u80fd\u4e0e\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u5bc6\u5207\u76f8\u5173\u3002\u901a\u8fc7\u805a\u7126\u8fd9\u4e9b\u5173\u952e\u7684\u51b3\u7b56\u70b9\uff0c\u6a21\u578b\u5b66\u5230\u7684\u53ef\u80fd\u662f\u66f4\u672c\u8d28\u7684\u63a8\u7406\u6a21\u5f0f\uff0c\u800c\u4e0d\u662f\u6b7b\u8bb0\u786c\u80cc\u7279\u5b9a\u9886\u57df\u7684\u77e5\u8bc6\u3002<\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"9\" data-puid=\"0\"><img decoding=\"async\" id=\"9\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/229\/w1080h749\/20250605\/991d-1f6e06488285017d416a989e05894d91.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u91cd\u65b0\u7406\u89e3\u5927\u6a21\u578b\u5f3a\u5316\u5b66\u4e60<\/p><p class=\"art_p\"><span>\u8fd9\u9879\u7814\u7a76\u7684\u610f\u4e49\u4e0d\u6b62\u4e8e\u63d0\u5347\u8bad\u7ec3\u6548\u7387\uff0c\u5728\u6df1\u5165\u7814\u7a76RLVR\uff08<\/span><span>Reinforcement Learning with Verifiable Rewards\uff09\u8bad\u7ec3\u8fc7\u7a0b\u540e\uff0c\u56e2\u961f\u53d1\u73b0\u4e86\u4e00\u4e2a\u4ee4\u4eba\u610f\u5916\u7684\u73b0\u8c61\uff1aRLVR\u5e76\u4e0d\u662f\u63a8\u5012\u91cd\u6765\uff0c\u800c\u662f\u5728base model\u7684\u57fa\u7840\u4e0a\u505a\u7cbe\u7ec6\u8c03\u6574\u3002<\/span><\/p><p class=\"art_p\">\u901a\u8fc7\u8ffd\u8e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u71b5\u6a21\u5f0f\u7684\u6f14\u53d8\uff0c\u53d1\u73b0\u5373\u4f7f\u5728\u8bad\u7ec3\u6536\u655b\u540e\uff08\u7b2c1360\u6b65\uff09\uff0c\u6a21\u578b\u4e0ebase model\u5728\u9ad8\u71b5token\u4f4d\u7f6e\u4e0a\u7684\u91cd\u53e0\u7387\u4ecd\u7136\u4fdd\u6301\u572886.67%\u4ee5\u4e0a\u3002\u8fd9\u610f\u5473\u7740RLVR\u57fa\u672c\u4fdd\u7559\u4e86\u9884\u8bad\u7ec3\u6a21\u578b\u5bf9\u201d\u54ea\u4e9b\u5730\u65b9\u8be5\u72b9\u8c6b\u3001\u54ea\u4e9b\u5730\u65b9\u8be5\u786e\u5b9a\u201d\u7684\u5224\u65ad\u3002<\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"10\" data-puid=\"0\"><img decoding=\"async\" id=\"10\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/496\/w1080h216\/20250605\/3dbf-82b422692b039c042dcb0fbf78ea03b1.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u66f4\u6709\u610f\u601d\u7684\u662f\uff0cRLVR\u7684\u8c03\u6574\u7b56\u7565\u76f8\u5f53\u201d\u504f\u5fc3\u201d\u3002\u56e2\u961f\u8ba1\u7b97\u4e86\u4e0d\u540c\u71b5\u6c34\u5e73token\u5728\u8bad\u7ec3\u524d\u540e\u7684\u53d8\u5316\uff0c\u53d1\u73b0\u4e86\u4e00\u4e2a\u6e05\u6670\u7684\u89c4\u5f8b\uff1a<\/p><p class=\"art_p\">\u521d\u59cb\u71b5\u8d8a\u9ad8\u7684token\uff0c\u5728RLVR\u8bad\u7ec3\u540e\u71b5\u7684\u589e\u5e45\u8d8a\u5927\u3002\u800c\u90a3\u4e9b\u539f\u672c\u5c31\u5f88\u786e\u5b9a\u7684\u4f4e\u71b5token\uff0c\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u51e0\u4e4e\u7eb9\u4e1d\u4e0d\u52a8\u3002<\/p><p class=\"art_p\">\u4ece\u71b5\u7684\u767e\u5206\u4f4d\u53d8\u5316\u56fe\u4e2d\u53ef\u4ee5\u770b\u5230\uff0c\u8d8a\u662f\u9ad8\u71b5\u7684token\uff08\u6bd4\u598290%\u300195%\u5206\u4f4d\uff09\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u6ce2\u52a8\u8303\u56f4\u8d8a\u5927\uff1b\u800c\u4f4e\u71b5token\uff08\u6bd4\u59820%\u30015%\u5206\u4f4d\uff09\u7684\u53d8\u5316\u5fae\u4e4e\u5176\u5fae\u3002<\/p><figure class=\"art_img_mini j_p_gallery\" data-type=\"simple\" data-startindex=\"11\" data-puid=\"0\"><img decoding=\"async\" id=\"11\" class=\"art_img_mini_img j_fullppt_cover finpic\" src=\"https:\/\/k.sinaimg.cn\/n\/sinakd20250605s\/581\/w1080h301\/20250605\/f8bf-4d0b1458ac65057a4bbbe1aaf0d52ff1.png\/w700d1q75cms.jpg?by=cms_fixed_width\" alt=\"\" \/><h2 class=\"art_img_tit\">\u00a0<\/h2><\/figure><p class=\"art_p\">\u6700\u540e\uff0c\u57fa\u4e8e\u4ee5\u4e0a\u6240\u6709\u8fd9\u4e9b\u53d1\u73b0\uff0c\u56e2\u961f\u8fd8\u63d0\u51fa\u4e86\u51e0\u4e2a\u6709\u8da3\u7684\u8ba8\u8bba\uff1a<\/p><p class=\"art_p\"><span>\u8ba8\u8bba1\uff1a<\/span>\u9ad8\u71b5token\u53ef\u80fd\u662f\u89e3\u91ca\u4e3a\u4ec0\u4e48\u5f3a\u5316\u5b66\u4e60\u80fd\u6cdb\u5316\u800c\u76d1\u7763\u5fae\u8c03\u503e\u5411\u4e8e\u8bb0\u5fc6\u800c\u8fc7\u62df\u5408\u7684\u5173\u952e\u3002<\/p><p class=\"art_p\">\u5f3a\u5316\u5b66\u4e60\u503e\u5411\u4e8e\u4fdd\u6301\u751a\u81f3\u589e\u52a0\u5206\u53c9token\u7684\u71b5\uff0c\u7ef4\u6301\u63a8\u7406\u8def\u5f84\u7684\u7075\u6d3b\u6027\u3002\u800c\u76d1\u7763\u5fae\u8c03\u4f1a\u628a\u8f93\u51falogits\u63a8\u5411one-hot\u5206\u5e03\uff0c\u5bfc\u81f4\u5206\u53c9token\u71b5\u964d\u4f4e\uff0c\u5931\u53bb\u4e86\u63a8\u7406\u8def\u5f84\u7684\u7075\u6d3b\u6027\u3002\u8fd9\u79cd\u7075\u6d3b\u6027\u53ef\u80fd\u6b63\u662f\u63a8\u7406\u6a21\u578b\u80fd\u591f\u6709\u6548\u6cdb\u5316\u5230\u672a\u89c1\u4efb\u52a1\u7684\u5173\u952e\u56e0\u7d20\u3002<\/p><p class=\"art_p\"><span>\u8ba8\u8bba2\uff1a\u00a0<\/span>\u4e0e\u4f20\u7edf\u5f3a\u5316\u5b66\u4e60\u4e0d\u540c\uff0c\u5927\u6a21\u578b\u63a8\u7406\u9700\u6574\u5408\u5148\u9a8c\u77e5\u8bc6\uff0c\u4e14\u5fc5\u987b\u751f\u6210\u53ef\u8bfb\u6027\u8f93\u51fa\u3002\u56e0\u6b64\uff0c\u5927\u6a21\u578b\u601d\u7ef4\u94fe\uff08CoT\uff09\u5305\u542b\u4f4e\u71b5\u7684\u591a\u6570\u6807\u8bb0\u4e0e\u9ad8\u71b5\u7684\u5c11\u6570\u6807\u8bb0\uff0c\u800c\u4f20\u7edf\u5f3a\u5316\u5b66\u4e60\u53ef\u5047\u8bbe\u6574\u4e2a\u8fc7\u7a0b\u4e2d\u52a8\u4f5c\u71b5\u662f\u5747\u5300\u5206\u5e03\u7684\u3002<\/p><p class=\"art_p\"><span>\u8ba8\u8bba3\uff1a<\/span>\u5728\u5f3a\u5316\u5b66\u4e60\u4e0e\u53ef\u9a8c\u8bc1\u5956\u52b1\u7b97\u6cd5\uff08RLVR\uff09\u4e2d\uff0c\u71b5\u5956\u52b1\uff08entropy bonus\uff09\u53ef\u80fd\u5e76\u975e\u6700\u4f18\u9009\u62e9\uff0c\u56e0\u4e3a\u5b83\u4f1a\u589e\u52a0\u4f4e\u71b5\u591a\u6570\u6807\u8bb0\u7684\u71b5\u503c\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cclip-higher\u65b9\u6cd5\u80fd\u6709\u6548\u63d0\u5347\u9ad8\u71b5\u5c11\u6570\u6807\u8bb0\u7684\u71b5\u503c\u3002<\/p><p class=\"art_p\">\u8bba\u6587<\/p><p class=\"art_p\">https:\/\/arxiv.org\/abs\/2506.01939<\/p><p class=\"art_p\">\u9879\u76ee\u4e3b\u9875<\/p><p class=\"art_p\">https:\/\/shenzhi-wang.github.io\/high-entropy-minority-tokens-rlvr\/<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t","protected":false},"excerpt":{"rendered":"<p>\u68a6\u6668 \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n","protected":false},"author":1,"featured_media":3890,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3888","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/3888","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/comments?post=3888"}],"version-history":[{"count":7,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/3888\/revisions"}],"predecessor-version":[{"id":3896,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/3888\/revisions\/3896"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media\/3890"}],"wp:attachment":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media?parent=3888"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/categories?post=3888"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/tags?post=3888"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}