<\/a>\u7ffb\u8bd1\u6280\u672f\u6559\u80b2\u4e0e\u7814\u7a76<\/span><\/span><\/p>\n

<\/p>\n

2024\u5e7410\u670816\u65e5 00:15<\/em>\u00a0<\/span>\u9655\u897f<\/span><\/em><\/span><\/section>\n

\n
$\"\"$ <\/p>\n
\u5f15\u8a00\uff1aEvaluation is All You Need<\/strong><\/span><\/section>\n
\u968f\u7740\u5bf9\u73b0\u6709\u4e92\u8054\u7f51\u6570\u636e\u7684\u9884\u8bad\u7ec3\u9010\u6e10\u6210\u719f\uff0c\u7814\u7a76\u7684\u63a2\u7d22\u7a7a\u95f4\u6b63\u7531\u9884\u8bad\u7ec3\u8f6c\u5411\u540e\u671f\u8bad\u7ec3\uff08Post-training\uff09<\/span>\uff0c<\/span>OpenAI o1\u7684\u53d1\u5e03\u6b63\u5f70\u663e\u4e86\u8fd9\u4e00\u70b9\u3002<\/span><\/section>\n
\u800cPost-training\u7684\u6838\u5fc3\u5728\u4e8e\u8bc4\u6d4b\uff08Evaluation\uff09\u3002\u53ef\u9760\u7684AI\u8bc4\u6d4b\u4e0d\u4ec5\u80fd\u5728\u590d\u6742\u4efb\u52a1\u7684\u8bc4\u6d4b\u4e2d\u63d0\u4f9b\u53ef\u6269\u5c55\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u51cf\u5c11\u4eba\u5de5\u52b3\u52a8\uff0c\u8fd8\u80fd\u5728\u5f3a\u5316\u5b66\u4e60\u4e2d\u751f\u6210\u6709\u6548\u7684\u5956\u52b1\u4fe1\u53f7\u5e76\u6307\u5bfc\u63a8\u7406\u8fc7\u7a0b\u3002<\/span><\/span><\/section>\n
\n
\n
\u4f8b\u5982\uff0c\u4e00\u4e2aAI\u8bc4\u6d4b\u5668\u53ef\u4ee5\u9075\u5faa\u7528\u6237\u8bbe\u8ba1\u7684\u8bc4\u5206\u6807\u51c6\uff0c\u5728\u89c6\u89c9\u5bf9\u8bdd\u4efb\u52a1\u4e2d\u4e3a\u4e0d\u540c\u6a21\u578b\u7684\u56de\u590d\uff08model response\uff09\u63d0\u4f9b1\u523010\u7684\u8bc4\u5206\u3002\u9664\u4e86\u8bc4\u5206\u5916\uff0c\u5b83\u8fd8\u4f1a\u63d0\u4f9b\u76f8\u5e94\u7684\u7ed9\u5206\u7406\u7531\uff0c\u786e\u4fdd\u6a21\u578b\u6027\u80fd\u8bc4\u6d4b\u7684\u900f\u660e\u6027\u548c\u4e00\u81f4\u6027\u3002<\/span><\/section>\n<\/section>\n<\/blockquote>\n
\u6765\u81ea\u5b57\u8282\u8df3\u52a8\u548c\u9a6c\u91cc\u5170\u5927\u5b66\u7684\u7814\u7a76\u56e2\u961f\u53d1\u5e03\u4e86\u9996\u4e2a\u7528\u4e8e\u591a\u4efb\u52a1\u8bc4\u6d4b\u7684<\/span>\u5f00\u6e90<\/span>\u591a\u6a21\u6001\u5927\u6a21\u578b<\/span>LLaVA-Critic<\/span><\/strong><\/span>\uff0c\u65e8\u5728\u542f\u53d1\u793e\u533a\u5f00\u53d1<\/span><\/span>\u901a\u7528\u5927\u6a21\u578b\u8bc4\u6d4b\u5668<\/strong><\/span>\uff08<\/span>generalist evaluator<\/span>\uff09<\/span><\/strong><\/span>\u3002<\/span><\/section>\n
$\"\"$ <\/p>\n
\n
\n
\u8bba\u6587\u6807\u9898\uff1aLLaVA-Critic: Learning to Evaluate Multimodal Models<\/span><\/p>\n<\/li>\n
\n
\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2410.02712<\/span><\/p>\n<\/li>\n
\n
\u9879\u76ee\u4e3b\u9875\uff1ahttps:\/\/llava-vl.github.io\/blog\/2024-10-03-llava-critic\/<\/span><\/p>\n<\/li>\n
\n
\u6570\u636e\u4e0e\u6a21\u578b\u5f00\u6e90\uff1ahttps:\/\/huggingface.co\/collections\/lmms-lab\/llava-critic-66fe3ef8c6e586d8435b4af8<\/span><\/p>\n<\/li>\n<\/ul>\n
<\/section>\n
\u9996\u5148\uff0c\u8be5\u56e2\u961f\u6784\u5efa\u4e86\u4e00\u4e2a\u6db5\u76d6\u4e86\u591a\u6837\u5316\u8bc4\u6d4b\u573a\u666f\u548c\u8bc4\u5206\u6807\u51c6\u7684\u8bc4\u6d4b\u6307\u4ee4\u9075\u5faa\u6570\u636e\u96c6\uff08critic instruction-following dataset\uff09\uff1b\u4e4b\u540e\uff0c\u5728\u8fd9\u4e00\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3LLaVA-Critic\uff0c\u4f7f\u4e4b\u5b66\u4f1a\u5bf9\u6a21\u578b\u56de\u590d\u7ed9\u51fa\u5408\u7406\u8bc4\u5206\uff08judgement\uff09\u548c\u6253\u5206\u4f9d\u636e\uff08reason\uff09\uff1b\u66f4\u8fdb\u4e00\u6b65\uff0c\u5728\u591a\u6a21\u6001\u8bc4\u6d4b\uff08LMM-as-a-Judge\uff09<\/strong><\/span>\u548c\u504f\u597d\u5b66\u4e60\uff08preference learning\uff09<\/span><\/strong>\u4e24\u4e2a\u573a\u666f\u4e2d\u9a8c\u8bc1\u4e86LLaVA-Critic\u7684\u6709\u6548\u6027\u3002<\/span><\/section>\n
\n
\n<\/section>\n
\u8bc4\u6d4b\u6307\u4ee4\u9075\u5faa\u6570\u636e\u96c6<\/strong><\/span><\/section>\n
\u8be5\u56e2\u961f\u9996\u5148\u6784\u5efa\u4e86\u4e00\u4e2a\u9ad8\u8d28\u91cf\u7684\u8bc4\u6d4b\u6307\u4ee4\u9075\u5faa\u6570\u636e\u96c6\uff0c\u65e8\u5728\u6db5\u76d6\u591a\u4e2a\u590d\u6742\u7684\u8bc4\u6d4b\u573a\u666f\uff0c\u6839\u636e\u5bf9\u5e94\u7684\u8bc4\u6d4b\u63d0\u793a\u7ed9\u51fa\u76f8\u5e94\u7684\u6253\u5206\u548c\u8bc4\u5206\u7406\u7531\u3002<\/span><\/section>\n
\u8bba\u6587\u4e2d\u5c06\u4f7f\u7528AI\u6a21\u578b\uff08\u5982GPT\uff09\u4f5c\u4e3a\u8bc4\u6d4b\u5668\u7684\u5f00\u653e\u5f0f\u591a\u6a21\u6001\u8bc4\u6d4b\u4efb\u52a1\u5206\u4e3a\u4e24\u7c7b\uff1a<\/span><\/section>\n
1. \u5355\u70b9\u8bc4\u5206\uff08pointwise-scoring\uff09\uff1a\u6839\u636e\u8bc4\u6d4b\u63d0\u793a\uff0c\u5bf9\u5355\u4e2a\u6a21\u578b\u56de\u590d\u8fdb\u884c\u6253\u5206\u3002<\/span><\/section>\n
2. \u6210\u5bf9\u6392\u5e8f\uff08pairwise-ranking\uff09\uff1a\u5bf9\u4e8e\u4e24\u4e2a\uff08\u4e00\u5bf9\uff09\u6a21\u578b\u56de\u590d\uff0c\u7ed9\u51fa\u4e8c\u8005\u4e4b\u95f4\u7684<\/span>\u504f\u5e8f\u5173\u7cfb<\/span>\u6216\u5ba3\u5e03\u5e73\u5c40\u3002<\/span><\/span><\/p>\n
$\"\"$ <\/a><\/p>\n
LLaVA-Critic-113k\u4e3b\u8981\u5305\u542b\u5355\u70b9\u8bc4\u5206\u548c\u6210\u5bf9\u6392\u5e8f\u4e24\u79cd\u8bc4\u6d4b\u8bbe\u5b9a\u3002\u5728\u4e24\u79cd\u8bc4\u6d4b\u4e2d\uff0cLLaVA-Critic\u5747\u9700\u8981\u6839\u636e\u7ed9\u5b9a\u7684\u56fe\u7247\u3001\u95ee\u9898\u3001\u6a21\u578b\u56de\u590d\u4ee5\u53ca\u8bc4\u6d4b\u63d0\u793a\u4e2d\u7ed9\u5b9a\u7684\u8bc4\u5206\u8981\u6c42\uff0c\u5bf9\u6a21\u578b\u56de\u590d\u6253\u5206\u5e76\u7ed9\u51fa\u7406\u7531\u3002<\/span><\/em><\/span><\/section>\n
\u9488\u5bf9\u5355\u70b9\u8bc4\u5206\uff0c\u8be5\u56e2\u961f\u4ece8\u4e2a\u591a\u6a21\u6001\u6570\u636e\u96c6\u4e2d\u6536\u96c6\u4e86\u8f93\u5165\u6307\u4ee4\uff08\u56fe\u7247-\u95ee\u9898\uff09\uff0c\u4f7f\u752813\u4e2aLMM\u751f\u6210\u6a21\u578b\u56de\u590d\uff0c\u5e76\u6c47\u96c6\u4e867\u4e2a\u5e38\u7528\u5f00\u653e\u5f0f\u8bc4\u6d4b\u57fa\u51c6\u4e2d\u7684\u8bc4\u6d4b\u63d0\u793a\uff0c\u7531\u6b64\u6574\u7406\u5f97\u5230\u8bc4\u6d4b\u6837\u672c\u3002\u9488\u5bf9\u4e8e\u6bcf\u4e00\u6761\u8bc4\u6d4b\u6837\u672c\uff0c\u518d\u8be2\u95eeGPT-4o\u8fdb\u884c\u8bc4\u6d4b\uff0c\u5f97\u5230\u5224\u65ad\u5f97\u5206\u4e0e\u7406\u7531\u3002<\/span><\/section>\n
\u9488\u5bf9\u6210\u5bf9\u6392\u5e8f\uff0c\u8be5\u56e2\u961f<\/span>\u6536\u96c6\u4e86\u4e09\u4e2a\u504f\u597d\u6570\u636e\u96c6\u4e2d\u7684\u6a21\u578b\u56de\u590d\uff0c\u8fd9\u4e9b\u6570\u636e\u4e2d\u5df2\u7ecf\u5305\u542b\u4e86\u4eba\u7c7b\u6216GPT-4V\u7684\u504f\u597d\u6392\u5e8f\u7ed3\u679c\u3002\u4e4b\u540e\uff0c\u5c06\u6bcf\u4e00\u5bf9\u6a21\u578b\u56de\u590d\u548c\u5df2\u77e5\u7684\u504f\u5e8f\u5173\u7cfb\u8f93\u5165\u7ed9GPT-4o\uff0c\u83b7\u53d6\u5176\u5bf9\u504f\u5e8f\u5173\u7cfb\u7684\u89e3\u91ca\u3002<\/span><\/section>\n
\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u4ed6\u4eec\u8bbe\u8ba1\u4e8630\u4e2a\u5305\u542b\u4e0d\u540c\u683c\u5f0f\u4e0e\u8bc4\u5206\u6807\u51c6\u7684\u8bc4\u6d4b\u63d0\u793a\u6a21\u677f\uff0c\u5c06\u3010\u56fe\u7247-\u95ee\u9898\u8f93\u5165\uff0c\u4e24\u4e2a\u6a21\u578b\u56de\u590d\uff0c\u504f\u5e8f\u5173\u7cfb\uff0c\u89e3\u91ca\u3011\u6253\u5305\u6210\u6db5\u76d6\u591a\u79cd\u8bc4\u6d4b\u573a\u666f\u7684\u8bc4\u6d4b\u6307\u4ee4\u9075\u5faa\u6570\u636e\u3002<\/span><\/section>\n
\u7531\u6b64\uff0cLLaVA-Critic-113k\u6570\u636e\u96c6\u5f97\u4ee5\u6784\u5efa\uff0c\u5171\u8ba1\u5305\u542b46k\u5f20\u56fe\u7247\u548c113k\u4e2a\u8bc4\u6d4b\u6570\u636e\u6837\u672c\u3002\u4e0b\u56fe\u5c55\u793a\u4e86\u5177\u4f53\u7684\u6570\u636e\u7edf\u8ba1\uff1a<\/span><\/section>\n
$\"\"$ <\/a><\/p>\n
LLaVA-Critic-113k\u6570\u636e\u96c6\u7684\u6570\u636e\u7edf\u8ba1\u3002\u8be5\u56e2\u961f\u6536\u96c6\u4e86\u5e7f\u6cdb\u7684\u8f93\u5165\u6307\u4ee4\u4e0e\u6a21\u578b\u56de\u590d\uff0c\u6db5\u76d6\u4e86\u591a\u4e2a\u8bc4\u6d4b\u4efb\u52a1\u548c\u9886\u57df\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u4ee5\u4e0a\u5168\u90e8\u6570\u636e\u6765\u6e90\u4e8e\u5f00\u6e90\u7684\u6307\u4ee4\u9075\u5faa\u8bad\u7ec3\u6570\u636e\uff0c\u4e0e\u5b9e\u9645\u7684\u8bc4\u6d4b\u57fa\u51c6\uff08evaluation benchmark\uff09\u6ca1\u6709\u91cd\u5408\u3002<\/span><\/em><\/span><\/section>\n
<\/section>\n
LLaVA-Critic<\/strong><\/span><\/section>\n
\u9996\u4e2a\u5177\u6709\u901a\u7528\u8bc4\u6d4b\u80fd\u529b\u7684<\/strong>\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b<\/strong><\/span><\/section>\n\u4e3a\u4e86\u4f7f\u6a21\u578b\u5177\u5907\u901a\u7528\u7684\u8bc4\u6d4b\u80fd\u529b\uff0c\u8be5\u56e2\u961f\u5bf9\u4e00\u4e2a\u5df2\u7ecf\u5177\u5907\u5f3a\u5927\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u7684\u9884\u8bad\u7ec3\u591a\u6a21\u6001\u5927\u6a21\u578b<\/strong><\/span>\u8fdb\u884c\u6307\u4ee4\u5fae\u8c03\u3002\u8fd9\u4e00\u70b9\u975e\u5e38\u5173\u952e\uff0c\u56e0\u4e3a\u6a21\u578b\u81ea\u8eab\u80fd\u9ad8\u8d28\u91cf\u5904\u7406\u590d\u6742\u89c6\u89c9\u4efb\u52a1\u662f\u5176\u5177\u5907\u8bc4\u6d4b\u80fd\u529b\u7684\u57fa\u7840\uff1b\u800c\u8bc4\u6d4b\u80fd\u529b\u5219\u5728\u6b64\u4e4b\u4e0a\uff0c\u4f5c\u4e3a\u9644\u52a0\u7684\u5224\u522b\u80fd\u529b\u5f97\u5230\u8fdb\u4e00\u6b65\u5f00\u53d1\u3002<\/span><\/section>\n
\u5728\u8bad\u7ec3\u4e2d\uff0cLLaVA-Critic\u4f1a\u63a5\u53d7\u4e00\u4e2a\u8bc4\u6d4b\u63d0\u793a\uff08evaluation prompt\uff09\uff0c\u5305\u542b\u591a\u6a21\u6001\u6307\u4ee4\u8f93\u5165\u3001\u6a21\u578b\u56de\u590d\u53ca\u53ef\u9009\u7684\u53c2\u8003\u56de\u590d\u3002\u5b83\u4f1a\u6839\u636e\u8bc4\u6d4b\u63d0\u793a\u4e2d\u7684\u8bc4\u5206\u6807\u51c6\uff0c\u9884\u6d4b\u5b9a\u91cf\u5206\u6570\u6216\u6210\u5bf9\u6392\u5e8f\uff0c\u5e76\u7ed9\u51fa\u8be6\u7ec6\u7684\u7406\u7531\u3002<\/span><\/section>\n
\u8be5\u56e2\u961f\u5bf9\u8bc4\u6d4b\u7ed3\u679c\uff08\u5206\u6570\u6216\u504f\u5e8f\u5173\u7cfb\uff09<\/span><\/strong>\u548c\u7406\u7531<\/strong><\/span>\u540c\u65f6\u5e94\u7528\u4ea4\u53c9\u71b5\u635f\u5931\u8fdb\u884c\u8bad\u7ec3\u3002\u5b9e\u9a8c\u4e2d\uff0c\u4ed6\u4eec\u4eceLLaVA-OneVision(OV)7B\/72B\u9884\u8bad\u7ec3\u6a21\u578b\u5f00\u59cb\uff0c\u4f7f\u7528LLaVA-Critic-113k \u6570\u636e\u96c6\u8fdb\u884c1\u8f6e\u5fae\u8c03\uff0c\u5f97\u5230LLaVA-Critic\u6a21\u578b\u3002<\/span><\/section>\n
\u573a\u666f\u4e00\uff1a\u591a\u6a21\u6001\u5927\u6a21\u578b\u4f5c\u4e3a\u8bc4\u6d4b\u5668\uff08LMM-as-a-Judge\uff09<\/strong><\/span><\/section>\n\u8be5\u56e2\u961f\u9996\u5148\u5bf9LLaVA-Critic\u6a21\u578b\u5728\u591a\u4e2a\u591a\u6a21\u6001\u8bc4\u6d4b\u4efb\u52a1\u4e2d\u4e0eGPT-4o\u4ee5\u53ca\u4eba\u7c7b\u7684\u6253\u5206\u4e00\u81f4\u6027\u8fdb\u884c\u4e86\u9a8c\u8bc1\u3002<\/span><\/section>\n
$\"\"$ <\/a><\/p>\n
\u5728\u8bc4\u5206\u7684\u6574\u4f53\u5206\u5e03\u548c\u5bf9\u56de\u590d\u6a21\u578b (response model) \u7684\u6392\u5e8f\u5c42\u9762\u4e0a\uff0cLLaVA-Critic\u5747\u5c55\u73b0\u4e86\u4e0eGPT-4o\u7684\u4e00\u81f4\u6027\u3002<\/span><\/em><\/span><\/p>\n
\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u5728\u5355\u70b9\u8bc4\u5206\u7684\u8bc4\u4f30\u4efb\u52a1\u4e2d\uff0cLLaVA-Critic\u5927\u5e45\u8d85\u8d8a\u5176\u57fa\u7840\u6a21\u578bLLaVA-OneVision\u2014\u2014\u5728\u591a\u4e2a\u5f00\u653e\u5f0f\u95ee\u7b54\u8bc4\u4f30\u57fa\u51c6\u4e0a\uff0c\u5176\u8bc4\u5206\u4e00\u81f4\u6027\u548c\u6a21\u578b\u6392\u5e8f\u4e0eGPT-4o\u9ad8\u5ea6\u543b\u5408\u3002<\/span><\/section>\n
$\"\"$ <\/a><\/p>\n
\u4e0a\u8868\u6bd4\u8f83\u4e86\u4e0d\u540c\u8bc4\u6d4b\u5668\u5728\u6210\u5bf9\u6392\u5e8f\u65b9\u9762\u4e0e\u4eba\u7c7b\u504f\u597d\u7684\u4e00\u81f4\u6027\u3002LLaVA-Critic-72B\u5728\u8bc4\u6d4b\u51c6\u786e\u7387\uff08\u4e0d\u5305\u542b\u5e73\u5c40\uff09\u65b9\u9762\u8fbe\u5230\u4e8673.6%\uff0c\u5df2\u7ecf\u8d85\u8d8a\u4e86GPT-4V\/4o\uff1b\u5728\u5176\u4f59\u4e24\u4e2a\u6307\u6807\u4e2d\u4e5f\u4e0e\u5546\u7528\u7684GPT\u6a21\u578b\u5dee\u5f02\u5f88\u5c0f\u3002<\/span><\/section>\n
\u5c3d\u7ba1LLaVA-Critic-7B\u7684\u6a21\u578b\u53c2\u6570\u5927\u5e45\u51cf\u5c11\uff0c\u4f46\u5176\u5728\u5305\u542b\u4e0e\u4e0d\u5305\u542b\u5e73\u5c40\u7684\u8bc4\u6d4b\u51c6\u786e\u7387\u4e0a\u4ecd\u5206\u522b\u8fbe\u5230\u4e8659.6% \u548c72.2%\u3002\u8fd9\u4e3a\u5728\u8d44\u6e90\u53d7\u9650\u7684\u73af\u5883\u4e2d\u90e8\u7f72LLaVA-Critic\u63d0\u4f9b\u4e86\u53ef\u884c\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/span><\/section>\n
$\"\"$ <\/a><\/p>\n
\u4e0a\u8868\u5c55\u793a\u4e86LLaVA-Critic\u5728MLLM-as-a-Judge\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u8868\u73b0\uff0c\u8be5\u6d4b\u8bd5\u5305\u62ec\u4e86\u8bad\u7ec3\u6570\u636e\u4e2d\u672a\u89c1\u8fc7\u7684\u66f4\u5e7f\u6cdb\u8bc4\u6d4b\u573a\u666f\u3002\u9762\u5bf9\u5168\u65b0\u7684\u8bc4\u6d4b\u4efb\u52a1\uff0cLLaVA-Critic\u4e5f\u663e\u8457\u7f29\u5c0f\u4e86\u5f00\u6e90\u6a21\u578b\u4e0eGPT-4o\/4V\u5728\u8bc4\u6d4b\u51c6\u786e\u6027\u4e0a\u7684\u5dee\u8ddd\uff0c\u5145\u5206\u5c55\u73b0\u5176\u6cdb\u5316\u6027\u4e0e\u901a\u7528\u6027\u3002<\/span><\/section>\n
\u5728\u4e0a\u8ff0\u5b9e\u9a8c\u7ed3\u679c\u4e2d\uff0c72B\u6a21\u578b\u7684\u8868\u73b0\u4f18\u4e8e7B\u6a21\u578b\uff0cCritic-7B\u7684\u8868\u73b0\u4e5f\u4f18\u4e8e\u4f7f\u7528\u5f31\u5316\u7248\u672c\u8bc4\u6d4b\u6570\u636e\u8bad\u7ec3\u7684Critic-7B\uff08v0.5\uff09\u2014\u2014 \u8fd9\u8fdb\u4e00\u6b65\u5f3a\u8c03\u4e86\u6a21\u578b\u6269\u5c55\uff08model scaling\uff09\u548c\u6570\u636e\u6269\u5c55\uff08data scaling\uff09\u5728\u6784\u5efa\u901a\u7528\u8bc4\u6d4b\u5668\u4e2d\u7684\u91cd\u8981\u6027\u3002<\/span><\/section>\n
$\"\"$ <\/a><\/p>\n
\u5728\u8fd9\u4e2a\u6210\u5bf9\u6392\u5e8f\u793a\u4f8b\u4e2d\uff0cLLaVA-Critic\u80fd\u591f\u51c6\u786e\u8bc6\u522b\u8f93\u5165\u56fe\u50cf\u7684\u5185\u5bb9\uff08\u624b\u5199\u6570\u5b57 \u300c7\u300d\uff09\uff0c\u5e76\u57fa\u4e8e\u56de\u590d\u95f4\u7684\u5dee\u5f02\u505a\u51fa\u5224\u65ad\uff0c\u7ed9\u51fa\u4e86\u548c\u4eba\u7c7b\u8bc4\u4f30\u8005\u4e00\u81f4\u7684\u6392\u5e8f\uff0c\u5e76\u63d0\u4f9b\u4e86\u6e05\u6670\u7684\u7406\u7531\u8bf4\u660e\u3002\u540e\u8005\uff08\u8bc4\u5206\u7406\u7531\uff09\u5bf9\u4e8e\u6784\u5efa\u53ef\u9760\u4eba\u5de5\u667a\u80fd\u81f3\u5173\u91cd\u8981\uff0c\u5b83\u4f7fLLaVA-Critic\u7684\u8bc4\u6d4b\u8fc7\u7a0b\u66f4\u900f\u660e\uff0c\u8bc4\u6d4b\u7ed3\u679c\u66f4\u53ef\u4fe1\u3002<\/span><\/section>\n
\u573a\u666f\u4e8c\uff1a\u504f\u597d\u5b66\u4e60\uff08Preference Learning\uff09<\/strong><\/span><\/section>\nLLaVA-Critic\u7684\u8bc4\u6d4b\u80fd\u529b\u4e5f\u53ef\u7528\u4e8e\u6bd4\u8f83\u6210\u5bf9\u6a21\u578b\u56de\u590d\u7684\u597d\u574f\uff0c\u4ece\u800c\u4f5c\u4e3a\u5956\u52b1\u4fe1\u53f7\u5e94\u7528\u4e8e<\/span>RLHF<\/span>\u548cDPO\u7b49\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u3002\u5b9e\u9a8c\u4e2d\uff0c\u8be5\u56e2\u961f\u5c06LLaVA-Critic\u7528\u4e8e\u8fed\u4ee3\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08iterative DPO\uff09\u7b97\u6cd5\uff0c\u5177\u4f53\u65b9\u5f0f\u5982\u4e0b\uff1a<\/span><\/span><\/section>\n
\n
\n
\n
\u7ed9\u5b9a\u4e00\u4e2a\u9884\u8bad\u7ec3LMM\u548c\u4e00\u7ec4\u56fe\u7247-\u95ee\u9898\u8f93\u5165\uff0c\u9996\u5148\u8ba9LMM\u5bf9\u6bcf\u4e00\u4e2a\u56fe\u7247-\u95ee\u9898\u8f93\u5165\u968f\u673a\u751f\u6210K=5\u4e2a\u5019\u9009\u56de\u590d\uff0c\u7531\u6b64\u6784\u5efa\u51faKx (K-1)=20\u4e2a\u6210\u5bf9\u56de\u590d\u3002<\/section>\n
\u63a5\u7740\uff0c\u4f7f\u7528LLaVA-Critic\u5bf9\u8fd920\u4e2a\u56de\u590d\u5bf9\u8fdb\u884c\u6210\u5bf9\u6392\u5e8f\uff0c\u9009\u51fa\u6700\u597d\u548c\u6700\u574f\u7684\u56de\u590d\uff0c\u5f62\u6210\u6210\u5bf9\u7684\u53cd\u9988\u6570\u636e\u96c6\u3002<\/section>\n
\u4e4b\u540e\uff0c\u4f7f\u7528\u8fd9\u4e00\u6570\u636e\u96c6\u5bf9\u4e8e\u9884\u8bad\u7ec3LMM\u8fdb\u884c\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff09\u8bad\u7ec3\u3002<\/section>\n
\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u6e10\u8fdb\u5f0f\u8fed\u4ee3\u8fd9\u4e00\u8fc7\u7a0b\u5171\u8ba1M\u8f6e\uff0c\u6bcf\u6b21\u4f7f\u7528\u6700\u65b0\u8bad\u7ec3\u7684\u6a21\u578b\u751f\u6210\u5019\u9009\u56de\u590d\uff0c\u6700\u7ec8\u5f97\u5230\u4e0eLLaVA-Critic\u53cd\u9988\u5bf9\u9f50\u7684\u6a21\u578b\u3002<\/section>\n<\/section>\n<\/section>\n<\/blockquote>\n
\u8be5\u56e2\u961f\u91c7\u7528LLaVA-OneVision\u4f5c\u4e3a\u521d\u59cbLMM\uff0c\u8fdb\u884c3\u8f6eiterative DPO\u8bad\u7ec3\uff0c\u6700\u7ec8\u5c06\u8bad\u7ec3\u540e\u7684\u6a21\u578b\u547d\u540d\u4e3aLLaVA-OneVision-Chat\u3002\u968f\u540e\uff0c\u4ed6\u4eec\u5728\u591a\u4e2a\u5f00\u653e\u5f0f\u95ee\u7b54\u8bc4\u6d4b\u57fa\u51c6\u4e0a\u6d4b\u8bd5\u4e86\u6700\u7ec8\u6a21\u578b\u7684\u8868\u73b0\uff0c\u4ee5\u6bd4\u8f83LLaVA-Critic\u548c\u5176\u4ed6\u5956\u52b1\u6a21\u578b\u7684\u6548\u679c\u3002<\/span><\/section>\n
\u5982\u4e0a\u8868\u6240\u793a\uff0c\u65e0\u8bba\u662f\u57287B\u8fd8\u662f72B\u57fa\u7840\u6a21\u578b\u4e0a\uff0cLLaVA-Critic\uff08AI\u53cd\u9988\uff09\u5747\u8d85\u8d8a\u4e86LLaVA-RLHF(\u4eba\u7c7b\u53cd\u9988)\uff0c\u663e\u8457\u63d0\u5347\u4e86\u57fa\u7840\u6a21\u578b\u57286\u4e2a\u591a\u6a21\u6001\u5f00\u653e\u5f0f\u95ee\u7b54\u8bc4\u6d4b\u57fa\u51c6\u4e0a\u7684\u8868\u73b0\u3002<\/span><\/section>\n
\u4e0b\u65b9\u7684\u67f1\u72b6\u56fe\u8fdb\u4e00\u6b65\u76f4\u89c2\u5c55\u793a\u4e86LLaVA-Critic\u7684\u53cd\u9988\u5bf9LLaVA-OneVision\u6a21\u578b\u5728\u89c6\u89c9\u95ee\u7b54\u6027\u80fd\u4e0a\u7684\u63d0\u5347\u6548\u679c\u3002\u53ef\u89c1\uff0cLLaVA-Critic\u4f5c\u4e3a\u4e00\u79cd\u63d0\u4f9b\u6709\u6548\u5956\u52b1\u4fe1\u53f7\u7684\u53ef\u6269\u5c55\u65b9\u6848\uff0c\u4e0d\u4ec5\u51cf\u5c11\u4e86\u5bf9\u6602\u8d35\u4eba\u5de5\u53cd\u9988\u7684\u4f9d\u8d56\uff0c\u8fd8\u901a\u8fc7AI\u751f\u6210\u7684\u53cd\u9988\u8fdb\u4e00\u6b65\u4f18\u5316\u4e86\u6a21\u578b\u7684\u504f\u597d\u5bf9\u9f50\u6548\u679c\u3002<\/span><\/section>\n
<\/section>\n
$\"\"$ <\/a><\/strong><\/span><\/section>\n
\u7ed3\u8bba<\/strong><\/span><\/section>\n
LLaVA-Critic\u662f\u9996\u4e2a\u901a\u7528\u7684\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b\u8bc4\u6d4b\u5668\uff0c\u80fd\u591f\u5728\u591a\u4e2a\u5f00\u653e\u5f0f\u591a\u6a21\u6001\u573a\u666f\u4e2d\u8bc4\u6d4b\u6a21\u578b\u8868\u73b0\u3002\u4e3a\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u7814\u7a76\u56e2\u961f\u7cbe\u5fc3\u6784\u5efa\u4e86\u4e00\u4e2a\u9ad8\u8d28\u91cf\u7684\u8bc4\u6d4b\u6307\u4ee4\u9075\u5faa\u6570\u636e\u96c6\uff0c\u6db5\u76d6\u591a\u6837\u5316\u7684\u8bc4\u6d4b\u4efb\u52a1\u4e0e\u6807\u51c6\u3002<\/span><\/section>\n
\u5b9e\u9a8c\u4e2d\u5c55\u793a\u4e86LLaVA-Critic\u5728\u4e24\u4e2a\u5173\u952e\u9886\u57df\u7684\u6709\u6548\u6027\uff1a<\/span><\/section>\n
1.\u00a0\u4f5c\u4e3a\u901a\u7528\u7684\u8bc4\u6d4b\u5668<\/span><\/strong>\uff0cLLaVA-Critic\u80fd\u591f\u4e3a\u9700\u8981\u8bc4\u6d4b\u7684\u6a21\u578b\u56de\u590d\u63d0\u4f9b\u5355\u70b9\u8bc4\u5206\u548c\u6210\u5bf9\u6392\u5e8f\uff0c\u8fd9\u4e9b\u8bc4\u5206\u548c\u6392\u5e8f\u4e0e\u4eba\u7c7b\u548cGPT-4o\u7684\u504f\u597d\u9ad8\u5ea6\u4e00\u81f4\uff0c\u4e3a\u81ea\u52a8\u8bc4\u6d4b\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u5f00\u653e\u5f0f\u56de\u590d\u63d0\u4f9b\u4e86\u4e00\u4e2a\u53ef\u884c\u7684\u5f00\u6e90\u66ff\u4ee3\u65b9\u6848\u3002<\/span><\/section>\n
2. \u5728\u504f\u597d\u5b66\u4e60<\/span><\/strong>\u65b9\u9762\uff0cLLaVA-Critic\u63d0\u4f9b\u7684\u504f\u597d\u4fe1\u53f7\u80fd\u6709\u6548\u63d0\u5347\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u89c6\u89c9\u5bf9\u8bdd\u80fd\u529b\uff0c\u751a\u81f3\u8d85\u8d8a\u4e86\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684LLaVA-RLHF\u5956\u52b1\u6a21\u578b\u3002<\/span><\/section>\n
\u8fd9\u9879\u5de5\u4f5c\u5728\u5229\u7528\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b\u81ea\u8eab\u7684\u8bc4\u4ef7\u80fd\u529b\u65b9\u9762\uff0c\u8fc8\u51fa\u4e86\u91cd\u8981\u7684\u4e00\u6b65\u3002\u6211\u4eec\u671f\u5f85\u66f4\u591a\u7684\u7814\u7a76\u53ef\u4ee5\u7531\u6b64\u51fa\u53d1\uff0c\u901a\u8fc7\u63a2\u7a76\u66f4\u5177\u53ef\u6269\u5c55\u6027\u7684\uff0c\u8d85\u8d8a\u4eba\u7c7b\u7684\u5bf9\u9f50\u53cd\u9988\u673a\u5236\uff0c\u8fdb\u4e00\u6b65\u63a8\u52a8\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u53d1\u5c55\u3002<\/span><\/section>\n
\n
\n
\u7279\u522b\u8bf4\u660e\uff1a\u672c\u6587\u4ec5\u7528\u4e8e\u5b66\u672f\u4ea4\u6d41\uff0c\u5982\u6709\u4fb5\u6743\u8bf7\u540e\u53f0\u8054\u7cfb\u5c0f\u7f16\u5220\u9664\u3002<\/span><\/span><\/section>\n<\/section>\n<\/blockquote>\n
<\/section>\n<\/section>\n","protected":false},"excerpt":{"rendered":"
\u6280\u672f\u79d1\u666e | Evaluation is All You Need\uff01\u9996\u4e2a\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b\u901a\u7528\u8bc4\u6d4b\u5668LLaVA- […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[82],"tags":[],"class_list":["post-35101","post","type-post","status-publish","format-standard","hentry","category-82"],"_links":{"self":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/35101","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=35101"}],"version-history":[{"count":1,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/35101\/revisions"}],"predecessor-version":[{"id":35118,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/35101\/revisions\/35118"}],"wp:attachment":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=35101"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=35101"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=35101"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}