{"id":1141033,"date":"2025-06-03T03:27:14","date_gmt":"2025-06-03T10:27:14","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&p=1141033"},"modified":"2025-08-21T23:54:13","modified_gmt":"2025-08-22T06:54:13","slug":"opa-dpo","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/opa-dpo\/","title":{"rendered":"OPA-DPO\uff1a\u591a\u6a21\u6001\u5927\u6a21\u578b\u5e7b\u89c9\u96be\u9898\u7684\u9ad8\u6548\u89e3\u51b3\u65b9\u6848"},"content":{"rendered":"\n
\u7f16\u8005\u6309\uff1a\u5728\u89c6\u89c9\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\u7684\u5feb\u901f\u53d1\u5c55\u4e2d\uff0c\u5e7b\u89c9\u95ee\u9898\u4e00\u76f4\u662f\u7814\u7a76\u8005\u4eec\u5173\u6ce8\u7684\u7126\u70b9\u3002\u6a21\u578b\u751f\u6210\u4e0e\u8f93\u5165\u56fe\u50cf\u4e0d\u4e00\u81f4\u751a\u81f3\u865a\u5047\u7684\u5185\u5bb9\uff0c\u4e0d\u4ec5\u5f71\u54cd\u7528\u6237\u4f53\u9a8c\uff0c\u4e5f\u963b\u788d\u4e86\u591a\u6a21\u6001\u6280\u672f\u5728\u5b9e\u9645\u573a\u666f\u4e2d\u7684\u843d\u5730\u3002\u5bf9\u6b64\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u548c\u9999\u6e2f\u4e2d\u6587\u5927\u5b66\u7684\u8054\u5408\u7814\u7a76\u56e2\u961f\u4ece\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff09\u5165\u624b\uff0c\u63d0\u51fa\u4e86 On-Policy Alignment (OPA)-DPO \u7b97\u6cd5\uff0c\u53ef\u901a\u8fc7\u786e\u4fdd\u8bad\u7ec3\u6570\u636e\u4e0e\u521d\u59cb\u7b56\u7565\uff08reference policy\uff09\u7684\u4e00\u81f4\u6027\uff0c\u6709\u6548\u89e3\u51b3\u5e7b\u89c9\u95ee\u9898\u3002\u8be5\u6210\u679c\u5df2\u83b7\u9009\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u9876\u4f1a CVPR 2025 \u7684 Oral \u8bba\u6587\u3002<\/p>\n\n\n\n
\u5728\u89c6\u89c9\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\u9886\u57df\uff0c\u751f\u6210\u4e0e\u8f93\u5165\u56fe\u50cf\u4e0d\u4e00\u81f4\u751a\u81f3\u8fd8\u6709\u865a\u5047\u5185\u5bb9\u7684\u201c\u5e7b\u89c9\u201d\u73b0\u8c61\uff0c\u662f\u4e00\u4e2a\u4e9f\u5f85\u653b\u514b\u7684\u6838\u5fc3\u96be\u9898\u3002\u4f5c\u4e3a\u4e00\u79cd\u7b80\u5355\u6709\u6548\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u76f4\u63a5\u504f\u597d\u4f18\u5316 (DPO) [1] \u6b63\u5728\u5f15\u8d77\u8d8a\u6765\u8d8a\u591a\u7684\u5173\u6ce8\u3002\u7814\u7a76\u8005\u4eec\u901a\u8fc7\u6bd4\u8f83\u6a21\u578b\u5728\u76f8\u540c\u63d0\u793a\u8bcd\u548c\u56fe\u50cf\u4e0b\u7684\u4e0d\u540c\u54cd\u5e94\uff0c\u6839\u636e\u5e7b\u89c9\u7a0b\u5ea6\u76f4\u63a5\u6784\u9020\u504f\u597d\u6570\u636e\u5bf9\uff0c\u7528\u4e8e DPO \u8bad\u7ec3\u3002<\/p>\n\n\n\n
\u7136\u800c\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u6ce8\u610f\u5230\uff0c\u73b0\u6709\u7814\u7a76\u4e2d\u4e0d\u540c\u7684\u6570\u636e\u6784\u9020\u65b9\u6cd5\u4f1a\u5bfc\u81f4\u663e\u8457\u7684\u6027\u80fd\u5dee\u5f02\u3002\u56e0\u6b64\uff0c\u4ed6\u4eec\u5bf9\u201c\u57fa\u4e8e DPO \u89e3\u51b3\u591a\u6a21\u6001\u5927\u6a21\u578b\u5e7b\u89c9\u95ee\u9898\u201d\u7684\u7b97\u6cd5\u8fdb\u884c\u4e86\u5168\u9762\u5206\u6790\uff0c\u603b\u7ed3\u4e86\u5b83\u4eec\u7684\u8868\u73b0\u53ca\u5c40\u9650\u6027\uff0c\u540c\u65f6\u4ece\u7406\u8bba\u89d2\u5ea6\u63ed\u793a\u4e86\u5404\u7b97\u6cd5\u6027\u80fd\u5dee\u5f02\u80cc\u540e\u7684\u6839\u672c\u539f\u56e0\uff0c\u5e76\u6307\u51fa\u51b3\u5b9a\u6a21\u578b\u6027\u80fd\u7684\u6700\u5173\u952e\u56e0\u7d20\u662f\u201c\u7528\u4e8e\u6784\u5efa\u504f\u597d\u5bf9\u7684\u6570\u636e\uff0c\u76f8\u8f83\u4e8e DPO \u5f00\u59cb\u524d\u7684\u7b56\u7565\uff08reference policy\uff09\u662f\u5426\u4e3a\u540c\u7b56\u7565\uff08on-policy\uff09\u201d\u3002<\/p>\n\n\n\n
\u7814\u7a76\u5458\u4eec\u5c06\u6b64\u524d\u7684\u7814\u7a76\u5de5\u4f5c\u5206\u4e3a\u4e09\u7c7b\uff1a<\/p>\n\n\n\n
\u7b2c\u4e00\u7c7b\u662f\u5e7b\u89c9\u6ce8\u5165\u7c7b\uff0c\u5982 HALVA [2]\u548c POVID [3]\uff0c\u901a\u8fc7\u5728\u5df2\u6709\u56fe\u50cf\u548c\u63d0\u793a\u7684\u6807\u51c6\u54cd\u5e94\u4e2d\u4eba\u4e3a\u6ce8\u5165\u5e7b\u89c9\u7247\u6bb5\u6765\u6784\u5efa\u504f\u597d\u5bf9\uff1b<\/p>\n\n\n\n
\u7b2c\u4e8c\u7c7b\u662f\u5e7b\u89c9\u8bc6\u522b\u7c7b\uff0c\u5982 RLHF-V [4]\u3001HA-DPO [5]\u548c HSA-DPO [6]\uff0c\u5148\u8ba9\u6a21\u578b\u6839\u636e\u56fe\u50cf\u548c\u63d0\u793a\u81ea\u884c\u751f\u6210\u54cd\u5e94\uff0c\u7136\u540e\u5229\u7528\u4e13\u5bb6\u53cd\u9988\uff08\u4eba\u7c7b\u6216 GPT-4\/4v\uff09\u6765\u8bc6\u522b\u548c\u4fee\u6539\u5176\u4e2d\u7684\u5e7b\u89c9\uff0c\u4ece\u800c\u6784\u5efa\u504f\u597d\u5bf9\uff1b<\/p>\n\n\n\n
\u7b2c\u4e09\u7c7b\u662f\u81ea\u6211\u8fdb\u5316\u7c7b\uff0c\u5982 RLAIF-V[7]\uff0c\u8ba9\u6a21\u578b\u9488\u5bf9\u540c\u4e00\u56fe\u50cf\u548c\u63d0\u793a\u751f\u6210\u591a\u4e2a\u54cd\u5e94\uff0c\u5e76\u7531\u4e00\u4e2a\u5728\u5e7b\u89c9\u8bc6\u522b\u65b9\u9762\u80fd\u529b\u66f4\u5f3a\u7684\u5bfc\u5e08\u6a21\u578b\u5bf9\u8fd9\u4e9b\u54cd\u5e94\u4e2d\u7684\u5e7b\u89c9\u4e25\u91cd\u7a0b\u5ea6\u8fdb\u884c\u5224\u65ad\u548c\u6392\u5e8f\uff0c\u4ee5\u6b64\u6784\u5efa\u504f\u597d\u5bf9\u3002<\/p>\n\n\n\n \u6839\u636e\u5b9e\u9a8c\u7ed3\u679c\uff0c\u8fd9\u4e09\u7c7b\u7b97\u6cd5\u7684\u6027\u80fd\u603b\u7ed3\u4e3a\uff1a\u81ea\u6211\u8fdb\u5316\u7c7b > \u5e7b\u89c9\u8bc6\u522b\u7c7b > \u5e7b\u89c9\u6ce8\u5165\u7c7b\u3002<\/p>\n\n\n\n \u5bf9\u4e8e\u5e7b\u89c9\u6ce8\u5165\u7c7b\uff0c\u5e7b\u89c9\u901a\u5e38\u5e76\u4e0d\u6765\u81ea\u6a21\u578b\u672c\u8eab\uff0c\u56e0\u6b64\u901a\u8fc7 DPO \u8bad\u7ec3\u5f80\u5f80\u4e0d\u80fd\u7ed9\u6a21\u578b\u5e26\u6765\u5f88\u5927\u589e\u76ca\u3002\u5bf9\u4e8e\u81ea\u6211\u8fdb\u5316\u7c7b\uff0c\u7406\u8bba\u4e0a\u7531\u4e8e\u7ef4\u5ea6\u707e\u96be\u95ee\u9898\uff0c\u8ba9\u6a21\u578b\u81ea\u884c\u63a2\u7d22\u5e76\u627e\u5230\u5b8c\u5168\u6b63\u786e\u7684\u56de\u590d\u662f\u5341\u5206\u56f0\u96be\u7684\uff0c\u6240\u4ee5\u90a3\u4e9b\u5b58\u5728\u4e8e\u591a\u4e2a\u56de\u590d\u4e2d\u7684\u987d\u56fa\u5e7b\u89c9\u901a\u5e38\u65e0\u6cd5\u901a\u8fc7\u8fd9\u79cd\u65b9\u6cd5\u6d88\u9664\u3002<\/p>\n\n\n\n \u76f4\u89c9\u4e0a\uff0c\u5e7b\u89c9\u8bc6\u522b\u7c7b\u7684\u65b9\u6cd5\u5e94\u8be5\u662f\u6700\u9ad8\u6548\u7684\u89e3\u51b3\u5e7b\u89c9\u7684\u65b9\u6848\uff0c\u90a3\u4e3a\u4ec0\u4e48\u5728\u5b9e\u8df5\u4e2d\u8fd9\u7c7b\u65b9\u6cd5\u5374\u8d25\u4e0b\u9635\u6765\uff1f\u4e3a\u4e86\u4e86\u89e3\u80cc\u540e\u7684\u539f\u56e0\uff0c\u7814\u7a76\u5458\u4eec\u4ece DPO \u7b97\u6cd5\u7684\u7ec6\u8282\u5165\u624b\u8fdb\u884c\u7814\u7a76\u3002<\/p>\n\n\n\n \u4e0e\u6700\u5e38\u7528\u7684 RLHF \u7b97\u6cd5 PPO \u7684\u521d\u59cb\u76ee\u6807\u76f8\u540c\uff0cDPO \u7684\u521d\u59cb\u76ee\u6807\u4e5f\u662f\uff08\u03c0_\u03b8 \u662f\u6a21\u578b\u7684\u5f53\u524d\u7b56\u7565\uff0c\u03c0_ref \u662f\u6a21\u578b\u7684\u521d\u59cb\u7b56\u7565\/\u53c2\u8003\u7b56\u7565\uff0cx \u4e3a\u63d0\u793a\u8bcd\uff0cm \u4e3a\u56fe\u50cf\uff0cy \u4e3a\u54cd\u5e94\uff0cr(x,y,m) \u662f\u901a\u8fc7 Bradley-Terry model \u8bad\u7ec3\u5f97\u5230\u7684\u5956\u52b1\u51fd\u6570\uff09\uff1a<\/p>\n\n\n\n \u5373\u5728\u6700\u5927\u5316\u5956\u52b1\u7684\u540c\u65f6\uff0c\u7ea6\u675f\u6a21\u578b\u5f53\u524d\u7b56\u7565\u4e0e\u6a21\u578b\u521d\u59cb\u7b56\u7565\u4e4b\u95f4\u7684 KL \u6563\u5ea6\u3002\u7136\u800c\uff0c\u7814\u7a76\u5458\u4eec\u91cd\u65b0\u5ba1\u89c6 KL \u6563\u5ea6\u7684\u5b9a\u4e49\u53d1\u73b0\uff0c\u7ed9\u5b9a\u4efb\u4f55\u4e00\u4e2a\u63d0\u793a\u8bcd\u548c\u56fe\u50cf (x,m)\uff0c\u82e5\u5b58\u5728\u4e00\u4e2a\u54cd\u5e94 (y) \u4f7f\u5f97 \u03c0_\u03b8(y|x,m)>0\uff0c\u4f46 \u03c0_ref(y|x,m)\u21920\uff0c\u6b64\u65f6 KL \u6563\u5ea6\u4f1a\u8d8b\u4e8e\u65e0\u7a77\u5927\u3002\u8fd9\u4e2a\u6027\u8d28\u8bf4\u660e\u2014\u2014\u5bf9\u4e8e\u4efb\u4f55\u4ece\u76ee\u6807\u51fd\u6570 (1) \u51fa\u53d1\u7684\u7b97\u6cd5\uff0c\u90a3\u4e9b\u76f8\u5bf9\u539f\u59cb\u7b56\u7565 (\u03c0_ref) \u91c7\u6837\u6982\u7387\u6781\u4f4e\u7684\u54cd\u5e94\uff08\u6839\u636e\u5f3a\u5316\u5b66\u4e60\u7684\u547d\u540d\u89c4\u8303\uff0c\u8fd9\u79cd\u6570\u636e\u88ab\u79f0\u4e3a\u5f02\u7b56\u7565\uff08off-policy\uff09\u6570\u636e\uff0c\u76f8\u53cd\u5219\u4e3a\u540c\u7b56\u7565\uff08on-policy\uff09\u6570\u636e\uff09\u5c06\u6ca1\u6709\u4efb\u4f55\u673a\u4f1a\u88ab\u6a21\u578b\u5b66\u4f1a\u3002<\/p>\n\n\n\n \u5982\u679c\u975e\u8981\u5c06\u8fd9\u4e9b\u5f02\u7b56\u7565\uff08off-policy\uff09\u7684\u4f18\u9009\u54cd\u5e94\uff08preferred response\uff09\u62ff\u6765\u6784\u5efa DPO \u504f\u597d\u5bf9\uff0c\u4f1a\u5bfc\u81f4\u68af\u5ea6\u5728\u4e0b\u4e00\u6b21\u66f4\u65b0\u65f6\u51e0\u4e4e\u6d88\u5931\u3002<\/p>\n\n\n\n \u91cd\u6e29 DPO \u8bad\u7ec3\u7684\u4f18\u5316\u76ee\u6807\uff1a<\/p>\n\n\n\n \u5176\u4e2d y_w \u662f\u4f18\u9009\u54cd\u5e94\uff08preferred response\uff09\uff0cy_l \u662f\u88ab\u62d2\u54cd\u5e94\uff08rejected response\uff09\uff0c\u5176\u68af\u5ea6\u53ef\u8868\u793a\u4e3a\uff08\u03c3(\u22c5) \u662f sigmoid \u51fd\u6570\uff09\uff1a<\/p>\n\n\n\n \u8bad\u7ec3\u5f00\u59cb\u524d \u03c0\u03b8=\u03c0_ref\uff0c\u6240\u4ee5 sigmoid \u51fd\u6570\u5185\u90e8\u7684\u503c\u5e94\u5f53\u4e3a0\uff0c\u5373\u5f53\u524d\u7b56\u7565\u4f1a\u4ee5 0.5\u03b2 \u4e3a\u7cfb\u6570\u5bf9 y_w \u8fdb\u884c\u6700\u5927\u5bf9\u6570\u4f3c\u7136\u66f4\u65b0\uff08max-loglikelihood update\uff09\u3002\u4f46\u662f\u5728\u8fd9\u4e00\u6b65\u66f4\u65b0\u8fc7\u540e\uff0clog\u03c0_ref(y_w\u2223x,m)\u03c0\u03b8(y_w\u2223x,m) \u5c06\u4f1a\u8d8b\u8fd1\u4e8e\u6781\u5927\u503c\uff08\u56e0\u4e3a\u5206\u5b50 > 0\uff0c\u800c\u5206\u6bcd\u8d8b\u8fd1\u4e8e0\uff09\uff0c\u4ece\u800c\u5bfc\u81f4 \u03c3(\u2212r_w+r_l)\u21920\u3002\u56e0\u6b64\uff0c\u68af\u5ea6\u4f1a\u5728\u4e0b\u4e00\u6b21\u66f4\u65b0\u65f6\u51e0\u4e4e\u6d88\u5931\u3002<\/p>\n\n\n\n \u56de\u987e\u5e7b\u89c9\u8bc6\u522b\u7c7b\u7684\u65b9\u6cd5\uff0c\u4e13\u5bb6\u6539\u52a8\u540e\u7684\u54cd\u5e94\uff0c\u5927\u90e8\u5206\u5bf9\u4e8e\u539f\u6a21\u578b\u6765\u8bf4\u90fd\u662f\u5f02\u7b56\u7565\uff08off-policy\uff09\u7684\uff0c\u5373\u4f7f\u8fd9\u4e9b\u6539\u52a8\u518d\u5fae\u5c0f\u4e5f\u65e0\u6d4e\u4e8e\u4e8b\uff0c\u6240\u4ee5\u6839\u672c\u65e0\u6cd5\u6307\u671b\u8fd9\u4e9b\u4e13\u5bb6\u53cd\u9988\u80fd\u88ab\u6a21\u578b\u5b66\u4f1a\u3002\u76f8\u5bf9\u5e94\u5730\uff0c\u81ea\u6211\u8fdb\u5316\u7c7b\u65b9\u6cd5\u5373\u4f7f\u5b58\u5728\u5b66\u4e60\u6548\u7387\u4e0d\u9ad8\u7684\u6f5c\u5728\u95ee\u9898\uff0c\u4f46\u662f\u5b83\u6784\u5efa\u7684\u504f\u597d\u5bf9\u90fd\u6765\u81ea\u6a21\u578b\u672c\u8eab\uff0c\u5373\u5168\u662f\u540c\u7b56\u7565\uff08on-policy\uff09\u7684\uff0c\u56e0\u6b64\u6548\u679c\u6700\u597d\u3002<\/p>\n\n\n\n \u662f\u5426\u5b58\u5728\u4e00\u79cd\u65b9\u6cd5\u65e2\u80fd\u591f\u5229\u7528\u4e13\u5bb6\u7684\u7cbe\u786e\u53cd\u9988\uff0c\u53c8\u80fd\u5b8c\u5168\u907f\u514d\u5f02\u7b56\u7565\uff08off-policy\uff09\u5bfc\u81f4\u7684 KL \u6563\u5ea6\u7ea6\u675f\u95ee\u9898\uff1f<\/p>\n\n\n\n \u9488\u5bf9\u73b0\u6709\u65b9\u6cd5\u7684\u5c40\u9650\u6027\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u8054\u5408\u9999\u6e2f\u4e2d\u6587\u5927\u5b66\u63d0\u51fa\u4e86\u4e00\u79cd\u7b80\u5355\u800c\u9ad8\u6548\u7684\u7b97\u6cd5 On-Policy Alignment(OPA)-DPO\uff0c\u5c06\u4e13\u5bb6\u7684\u7cbe\u786e\u53cd\u9988\u6570\u636e\u5728 DPO \u8bad\u7ec3\u524d\u4e0e\u6a21\u578b\u7b56\u7565\u5bf9\u9f50\u3002\u5728\u4ec5\u4f7f\u75284.8k\u6570\u636e\u7684\u60c5\u51b5\u4e0b\uff0cOPA-DPO \u53ef\u4ee5\u5b9e\u73b0\u76ee\u524d SOTA \u7684\u6027\u80fd\uff0c\u800c\u4e4b\u524d\u7684 SOTA \u7b97\u6cd5\u9700\u898116k\u6570\u636e\u3002\u8be5\u6210\u679c\u5df2\u83b7\u9009\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u9876\u4f1a CVPR 2025 \u7684 Oral \u8bba\u6587\u3002<\/p>\n\n\n\n Mitigating Hallucinations in Large Vision-Language Models via DPO: On-Policy Data Hold the Key<\/p>\n\n\n\n
<\/figure>\n\n\n\n
<\/figure>\n\n\n\n
<\/figure>\n\n\n\n
<\/figure>\n\n\n\nOPA-DPO\uff1a\u6253\u7834\u5e38\u89c4\uff0c\u91cd\u5851\u5bf9\u9f50\u7b56\u7565<\/h2>\n\n\n\n