{"id":14235,"date":"2024-11-27T19:49:23","date_gmt":"2024-11-27T11:49:23","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=14235"},"modified":"2024-11-27T19:49:23","modified_gmt":"2024-11-27T11:49:23","slug":"cong-openai-o1-kanda","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/cong-openai-o1-kanda\/","title":{"rendered":"\u4ece OpenAI-o1 \u770b\u5927\u6a21\u578b\u7684\u590d\u6742\u63a8\u7406\u80fd\u529b"},"content":{"rendered":"

2022 \u5e74 OpenAI \u53d1\u5e03\u4e86 ChatGPT\uff0c\u6210\u4e3a\u4e16\u754c\u4e0a\u6700\u5feb\u7a81\u7834\u4e0a\u4ebf\u7528\u6237\u7684 APP\uff0c\u90a3\u65f6\u5019\u4eba\u4eec\u90fd\u8ba4\u4e3a\uff0c\u6211\u4eec\u79bb\u771f\u6b63\u7684\u4eba\u5de5\u667a\u80fd\u66f4\u8fd1\u4e86\u3002\u4f46\u662f\u4eba\u4eec\u5f88\u5feb\u53d1\u73b0\uff0cChatGPT \u53ef\u4ee5\u5bf9\u8bdd\u804a\u5929\uff0c\u751a\u81f3\u53ef\u4ee5\u5199\u8bd7\u5199\u6587\u7ae0\uff0c\u4f46\u5728\u7b80\u5355\u7684\u903b\u8f91\u4e0a\u4f9d\u7136\u4e0d\u5c3d\u5982\u4eba\u610f\uff0c\u6bd4\u5982\u90a3\u4e2a\u8457\u540d\u7684 “strawberry” \u4e2d\u6709\u51e0\u4e2a “r” \u7684\u6897\u3002<\/p>\n

\u800c\u4e24\u5e74\u4ee5\u540e\u7684\u4eca\u5929\uff0cOpenAI \u53c8\u53d1\u5e03\u4e86 o1 \u6a21\u578b\uff0c\u4ee5\u5176\u5f3a\u5927\u7684\u903b\u8f91\u63a8\u7406\u80fd\u529b\u548c OpenAI \u5f3a\u5927\u7684\u6280\u672f\u9690\u85cf\u80fd\u529b\u5f15\u53d1\u4e86\u4eba\u4eec\u5bf9\u5b83\u80cc\u540e\u65b9\u6cd5\u7684\u70ed\u70c8\u8ba8\u8bba\u3002\u672c\u6587\u68b3\u7406\u4e86\u4e00\u4e9b\u76f8\u5173\u6587\u7ae0\uff0c\u4ee5\u5bf9 o1 \u6a21\u578b\u6280\u672f\u7684\u731c\u6d4b\u4e3a\u5f15\uff0c\u770b\u4e00\u770b\u5927\u6a21\u578b\u7684\u590d\u6742\u63a8\u7406\u80fd\u529b\u7684\u53d1\u5c55\u5386\u7a0b\u3002<\/p>\n

 <\/p>\n

01 \u80cc\u666f\u4ecb\u7ecd<\/h2>\n

 <\/p>\n

\u601d\u7ef4\u94feChain of Thought\uff0c\u7b80\u79f0 CoT\uff0c\u662f\u4e00\u79cd\u8ba4\u77e5\u5fc3\u7406\u5b66\u548c\u6559\u80b2\u5b66\u4e2d\u7684\u6982\u5ff5\uff0c\u5b83\u63cf\u8ff0\u4e86\u4eba\u4eec\u5728\u89e3\u51b3\u95ee\u9898\u6216\u505a\u51fa\u51b3\u7b56\u65f6\u601d\u7ef4\u7684\u9010\u6b65\u53d1\u5c55\u8fc7\u7a0b\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u4e0d\u4ec5\u4ec5\u662f\u7b80\u5355\u5730\u4ece\u95ee\u9898\u76f4\u63a5\u8df3\u8dc3\u5230\u7b54\u6848\uff0c\u800c\u662f\u5305\u542b\u4e86\u591a\u4e2a\u6b65\u9aa4\uff0c\u6bcf\u4e2a\u6b65\u9aa4\u90fd\u53ef\u80fd\u6d89\u53ca\u4fe1\u606f\u7684\u6536\u96c6\u3001\u5206\u6790\u3001\u8bc4\u4f30\u4ee5\u53ca\u5bf9\u5148\u524d\u7ed3\u8bba\u7684\u4fee\u6b63\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u4e2a\u4f53\u80fd\u591f\u66f4\u7cfb\u7edf\u5730\u5904\u7406\u590d\u6742\u7684\u95ee\u9898\uff0c\u5e76\u6784\u5efa\u51fa\u5408\u7406\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n

\u76d1\u7763\u5fae\u8c03Supervised Learning\uff0c\u76d1\u7763\u5b66\u4e60\uff0c\u662f\u673a\u5668\u5b66\u4e60\u9886\u57df\u6700\u5e38\u89c1\u7684\u6a21\u578b\u8bad\u7ec3\u65b9\u5f0f\uff0c\u4f7f\u7528\u6807\u6ce8\u597d\u7684\u6570\u636e\u96c6\u8ba9\u6a21\u578b\u4ece\u4e2d\u5b66\u4e60\uff0c\u4ee5\u51c6\u786e\u5206\u7c7b\u6570\u636e\u6216\u9884\u6d4b\u7ed3\u679c\u3002\u8f93\u5165\u6570\u636e\u8fdb\u5165\u5230\u6a21\u578b\u4e2d\u65f6\uff0c\u76d1\u7763\u5b66\u4e60\u4f1a\u8c03\u6574\u6a21\u578b\u7684\u6743\u91cd\uff0c\u76f4\u5230\u6a21\u578b\u4ea7\u751f\u9002\u5f53\u7684\u62df\u5408\u3002
\nSupervised Fine-Tune\uff0c\u7b80\u79f0 SFT\uff0c\u662f\u6307\u6211\u4eec\u5728\u5df2\u6709\u57fa\u5ea7\u6a21\u578b\u7684\u4e0a\uff0c\u4f7f\u7528\u4e13\u6ce8\u4e8e\u67d0\u4e00\u9879\u7279\u6b8a\u4efb\u52a1\u7684\u6570\u636e\u96c6\u8bad\u7ec3\u6a21\u578b\u8fdb\u884c\u76d1\u7763\u5b66\u4e60\uff0c\u4ee5\u671f\u5176\u80fd\u4ece\u4e2d\u5b66\u4e60\u5230\u89e3\u51b3\u7279\u6b8a\u4efb\u52a1\u7684\u80fd\u529b\u3002<\/p>\n

\u5f3a\u5316\u5b66\u4e60Reinforcement Learning\uff0c\u7b80\u79f0 RL\uff0c\u5f3a\u5316\u5b66\u4e60\u662f\u4e09\u79cd\u57fa\u672c\u673a\u5668\u5b66\u4e60\u8303\u5f0f\u4e4b\u4e00\uff0c\u4e0e\u76d1\u7763\u5b66\u4e60\u548c\u65e0\u76d1\u7763\u5b66\u4e60\u5e76\u5217\u3002\u5f3a\u5316\u5b66\u4e60\u7684\u91cd\u70b9\u662f\u5728\u63a2\u7d22\uff08\u672a\u77e5\uff09\u548c\u5229\u7528\uff08\u5df2\u77e5\uff09\u4e4b\u95f4\u627e\u5230\u5e73\u8861\uff0c\u8ba9\u6a21\u578b\u4ee5\u6700\u5927\u5316\u957f\u671f\u56de\u62a5\u4e3a\u76ee\u6807\uff0c\u5b66\u4e60\u5230\u6b63\u786e\u7684\u884c\u4e3a\u3002
\n<\/p>\n

\u56fe\u7247\u6765\u81eaAWS\uff0c\u5982\u56fe\u6240\u793a\uff0c\u5728\u5f3a\u5316\u5b66\u4e60\u4e2d\uff0c\u667a\u80fd\u4f53 (Agent) \u662f\u6211\u4eec\u6700\u7ec8\u9700\u8981\u8bad\u7ec3\u7684\u76ee\u6807\uff0c\u4e0e\u8bbe\u5b9a\u597d\u7684\u73af\u5883 (Environment) \u8fdb\u884c\u4ea4\u4e92\u884c\u4e3a (Action) \u5e76\u4ea7\u751f\u5956\u52b1 (Reward) \u548c\u72b6\u6001 (state) \u8f6c\u79fb\uff0cAgent \u6839\u636e\u5956\u52b1\u8fdb\u884c\u5b66\u4e60\u66f4\u597d\u5730\u9009\u62e9\u4e0b\u4e00\u6b65 Action\uff0c\u5982\u6b64\u5faa\u73af\u5c31\u662f\u5f3a\u5316\u5b66\u4e60\u7684\u8bad\u7ec3\u8fc7\u7a0b\u3002<\/p>\n

\u5728 LLM \u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0cRL \u8d77\u4e86\u91cd\u8981\u4f5c\u7528\uff0c\u9884\u8bad\u7ec3\u9636\u6bb5\u501f\u52a9 RLHF \u8fdb\u884c\u5bf9\u9f50\u5df2\u7ecf\u6210\u4e86\u4e1a\u754c\u5171\u8bc6\u3002\u5728 LLM \u7684\u5f3a\u5316\u5b66\u4e60\u4e2d\uff0c\u6211\u4eec\u901a\u5e38\u9700\u8981\u4e00\u4e2a\u53e6\u4e00\u4e2a\u6a21\u578b\u6a21\u62df\u73af\u5883\uff0c\u5bf9 LLM \u7684\u8f93\u51fa\u8fdb\u884c\u5956\u52b1\uff0c\u8fd9\u4e2a\u6a21\u578b\u79f0\u4e3a Reward Model\uff0c\u7b80\u79f0 RM\u3002<\/p>\n

\u6211\u4eec\u4f1a\u6709\u591a\u4e2a\u6a21\u578b\uff1aActor \u6a21\u578b\u3001Critic \u6a21\u578b\u548c Reward \u6a21\u578b\u3002\u4e0e\u4e0a\u6587\u6807\u51c6 RL \u7684\u8bad\u7ec3\u6846\u67b6\u4e00\u81f4\uff0c\u5728 RL \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u7531 Actor \u548c Critic \u7ec4\u6210 Agent\uff0cReward \u5219\u4f5c\u4e3a Environment \u8fdb\u884c\u8bad\u7ec3\u3002
\n<\/p>\n

\u4f46\u662f\u8bad\u7ec3\u7ed3\u675f\u4e4b\u540e\uff0c\u6211\u4eec\u53ef\u4ee5\u5355\u72ec\u90e8\u7f72 Actor \u6216\u8005 Reward \u6a21\u578b\uff0c\u5176\u4e2d Actor \u6a21\u578b\u662f\u6211\u4eec\u7684 Generator\uff0cReward \u6a21\u578b\u5219\u662f\u6211\u4eec\u7528\u4e8e\u8861\u91cf Generator \u751f\u6210\u8d28\u91cf\u7684 Verifier\uff0c\u8fd9\u5c31\u662f OpenAI \u5728 Let’s verify step by step \u8fd9\u7bc7\u8bba\u6587\u4e2d\u63d0\u5230\u7684 Generator-Verifier \u7ed3\u6784\u3002<\/p>\n

\u800c Reward Model \u53ef\u4ee5\u6839\u636e\u5176\u53cd\u9988\u7684\u7ec6\u81f4\u7a0b\u5ea6\u5206\u4e3a\uff1a<\/p>\n

-\u57fa\u4e8e\u8fc7\u7a0b\u7684\u5956\u52b1\u6a21\u578b PRM\uff1aPRM \u4f1a\u6839\u636e LLM \u7684\u4e2d\u95f4\u7ed3\u679c\u7ed9\u4e88\u53cd\u9988\u3002
\n-\u57fa\u4e8e\u7ed3\u679c\u7684\u5956\u52b1\u6a21\u578b ORM\uff1aORM \u5219\u662f\u5728\u6700\u7ec8\u7ed3\u679c\u4e4b\u540e\u624d\u7ed9\u4e88\u53cd\u9988\u3002
\n\u5173\u4e8e\u8fd9\u4e24\u4e2a\u6982\u5ff5\u6211\u4eec\u5728\u4e0b\u6587\u7684\u5177\u4f53\u573a\u666f\u4e2d\u4ecb\u7ecd\u3002<\/p>\n

\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22Monte Carlo Tree Search\uff0c\u7b80\u79f0 MCTS\uff0c\u662f\u4e00\u4e2a\u6811\u641c\u7d22\u7b97\u6cd5\uff0c\u6838\u5fc3\u601d\u60f3\u662f\uff0c\u6bcf\u4e00\u6b65\u5c1d\u8bd5\u591a\u4e2a\u884c\u4e3a\u5e76\u9884\u6d4b\u884c\u4e3a\u672a\u6765\u7684\u53ef\u80fd\u6536\u76ca\uff0c\u4fa7\u91cd\u4e8e\u9009\u62e9\u6027\u63a2\u7d22\u4e00\u4e9b\u66f4\u5177\u6709\u6536\u76ca\u7684\u884c\u4e3a\u3002
\n
\n\u56fe\u7247\u6765\u81ea\u7ef4\u57fa\u767e\u79d1\u3002\u6bcf\u4e00\u6b21\u63a2\u7d22\u79f0\u4e3a\u5206\u4e3a\u56db\u4e2a\u6b65\u9aa4\uff1a<\/p>\n

-Selection\uff1a\u9009\u62e9\u4e00\u4e2a\u7ed3\u70b9
\n-Expansion\uff1a\u4ece\u8fd9\u4e2a\u7ed3\u70b9\u751f\u6210\u4e00\u4e2a\u65b0\u7684\u7ed3\u70b9\u8fdb\u884c\u63a2\u7d22
\n-Rollout\uff1a\u6cbf\u7740\u8fd9\u4e2a\u65b0\u7684\u7ed3\u70b9\u8fdb\u884c\u4e00\u6b21\u6a21\u62df\uff0c\u4ea7\u751f\u4e00\u4e2a\u7ed3\u679c
\n-Back Propagation\uff1a\u6a21\u62df\u7684\u7ed3\u679c\u53cd\u5411\u4f20\u64ad\uff0c\u66f4\u65b0\u8def\u5f84\u4e0a\u7684\u7ed3\u70b9
\n\u901a\u8fc7\u4e0d\u65ad\u7684\u63a2\u7d22\uff0c\u6211\u4eec\u5c31\u80fd\u5f97\u5230\u4e00\u68f5\u6811\uff0c\u5e76\u4e14\u6bcf\u4e00\u4e2a\u7ed3\u70b9\u90fd\u6709\u63a2\u7d22\u7684\u53ef\u80fd\u7ed3\u679c\uff0c\u6211\u4eec\u53ef\u4ee5\u5728\u8fd9\u68f5\u6811\u4e2d\u641c\u7d22\u5f97\u5230\u6700\u4f73\u7684\u8def\u5f84\u6216\u7ed3\u679c\u3002<\/p>\n

MCTS \u7528\u4e8e RL \u65f6\u4ea7\u751f\u8fc7 AlphaZero \u8fd9\u6837\u77e5\u540d\u7684\u6a21\u578b\u3002AlphaZero \u7684\u505a\u6cd5\u662f\u4f7f\u7528\u53ef\u8bad\u7ec3\u7684\u6a21\u578b\u6267\u884c Selection \u548c Rollout \u7684\u6b65\u9aa4\u3002\u4ece\u800c\u51cf\u5c11 MCTS \u5e9e\u5927\u7684\u641c\u7d22\u7a7a\u95f4\u548c\u6a21\u62df\u6210\u672c\uff0c\u4ee5\u9ad8\u6548\u7684\u5f97\u5230\u6700\u4f18\u89e3\uff0c\u6bd4\u5982\uff1a\u4f7f\u7528 Policy Network \u9ad8\u6548\u641c\u7d22\u4e0b\u4e00\u6b65\u7684\u53ef\u80fd\uff1b\u4f7f\u7528 Value Network \u5224\u65ad\u6bcf\u4e00\u6b65\u7684\u4ef7\u503c\u800c\u975e Rollout \u6a21\u62df\u3002<\/p>\n

o1 \u7684\u591a\u6b65\u63a8\u7406\u80fd\u529b\u63d0\u5230 o1 \u6a21\u578b\uff0c\u5c31\u4e0d\u5f97\u4e0d\u8bf4\u8d77\u4ee4\u4eba\u60ca\u8bb6\u7684\u591a\u6b65\u63a8\u7406\u80fd\u529b\u3002OpenAI \u5b98\u7f51\u7ed9\u51fa\u7684\u51e0\u4e2a\u4f8b\u5b50\u76f4\u89c2\u7684\u5c55\u793a\u4e86\u5176\u5728 \u5bc6\u7801\u3001\u4ee3\u7801\u3001\u6570\u5b66\u3001\u586b\u5b57\u6e38\u620f\u7b49\u591a\u4e2a\u65b9\u9762\u7684\u591a\u6b65\u63a8\u7406\u80fd\u529b\u3002\u5176\u4e2d\u201c\u5bc6\u7801\u201d\u76f8\u5173\u7684\u793a\u4f8b\uff0c\u89e3\u7801\u5f97\u5230\u7684\u7ed3\u679c\u662f \u201cTHERE ARE THREE R\u2019S IN STRAWBERRY\u201d \u4e5f\u662f\u5bf9\u66fe\u7ecf ChatGPT<\/a> \u63a8\u7406\u80fd\u529b\u7684\u56de\u5e94\u3002<\/p>\n

<\/p>\n

\u56e0\u6b64\uff0c\u6211\u4eec\u4e3b\u8981\u63a2\u7a76\u4e86\u5728\u8fd9\u9879\u80fd\u529b\u65b9\u9762\u7684\u4e00\u4e9b\u8bba\u6587\uff0c\u6574\u7406\u603b\u7ed3\u5982\u4e0b\u6587\u6240\u8ff0\u3002<\/p>\n

 <\/p>\n

02 \u63d0\u793a\u8bcd\u5de5\u7a0b<\/h2>\n

 <\/p>\n

\u5728\u4ecb\u7ecd\u7528\u4e8e\u63d0\u5347\u6a21\u578b\u63a8\u7406\u80fd\u529b\u7684\u63d0\u793a\u8bcd\u5de5\u7a0b\u4e4b\u524d\uff0c\u6211\u4eec\u5148\u8981\u4e86\u89e3\uff0c\u4ec0\u4e48\u662f Few-Shot Learning\u3002\u76ee\u524d\u4eba\u5de5\u667a\u80fd\u7684\u8bad\u7ec3\u666e\u904d\u9700\u8981\u5927\u91cf\u7684\u793a\u4f8b\u6570\u636e\uff0c\u800c\u5982\u679c\u53ea\u4f7f\u7528\u5f88\u5c11\u7684\u793a\u4f8b\u6570\u636e\u8fdb\u884c\u5b66\u4e60\uff0c\u5c31\u88ab\u79f0\u4e3a Few-Shot\uff0c\u5982\u679c\u5b8c\u5168\u4e0d\u7ed9\u793a\u4f8b\uff0c\u5219\u79f0\u4e3a Zero-Shot\u3002
\n\u300aChain of Thought Prompting Elicits Reasoning in Large Language Models\u300b\u8fd9\u7bc7\u8bba\u6587\u63d0\u51fa\u7684\u5c31\u662f\u4e00\u79cd Few-Shot \u7684\u65b9\u6cd5\u63d0\u5347\u6a21\u578b\u7684\u6570\u5b66\u63a8\u7406\u80fd\u529b\uff1a<\/p>\n

<\/p>\n

\u5982\u56fe\u6240\u793a\uff0c\u5de6\u4fa7\u5728\u8f93\u5165 LLM \u7684 Prompt \u4e2d\uff0c\u7ed9\u51fa\u4e86\u4e00\u4e2a\u6837\u4f8b\u8ba9 LLM \u5b66\u4e60\uff0c\u8fd9\u5c31\u662f Few-Shot Learning\uff0c\u4f46\u662f\u5176\u6548\u679c\u4f9d\u7136\u4e0d\u7406\u60f3\u3002\u8bba\u6587\u63d0\u51fa\u4e86\u53f3\u8fb9\u8fd9\u79cd\u5e26\u6709 CoT \u7684 Few-Shot \u8303\u5f0f\u3002\u6240\u4ee5\u53f3\u4fa7\u5728 Few-Shot \u4e2d\uff0c\u4e0d\u4ec5\u7ed9\u51fa\u4e00\u4e2a\u793a\u4f8b\u7684\u95ee\u9898\u548c\u7b54\u6848\uff0c\u8fd8\u7ed9\u51fa\u4e86\u4e2d\u95f4\u8fc7\u7a0b\u548c\u7ed3\u679c\u3002\u4f5c\u8005\u53d1\u73b0\uff0c\u8fd9\u6837\u4f7f\u7528 CoT \u6784\u9020\u7684 Few-Shot Prompt \u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n

\u800c\u968f\u7740\u6a21\u578b\u672c\u8eab\u80fd\u529b\u7684\u63d0\u9ad8\u4ee5\u53ca\u7814\u7a76\u7684\u6df1\u5165\uff0c\u300aLarge Language Models are Zero-Shot Reasoners\u300b\u8fd9\u7bc7\u6587\u7ae0\u8fdb\u4e00\u6b65\u53d1\u73b0 Zero-Shot \u4e5f\u53ef\u4ee5\u4f7f\u7528 CoT \u6765\u589e\u5f3a\u6a21\u578b\u80fd\u529b\u4e86\uff1a<\/p>\n

<\/p>\n

\u65e0\u9700\u8d39\u5fc3\u8d39\u529b\u6784\u9020\u4e00\u4e2a CoT \u4e2d\u95f4\u8fc7\u7a0b\uff0c\u751a\u81f3\u65e0\u9700\u6784\u9020\u793a\u4f8b\u8fdb\u884c Few-Shot\uff0c\u4e00\u53e5\u7b80\u5355\u7684 \u201cLet’s think step by step\u201d \u5c31\u53ef\u4ee5\u589e\u5f3a LLM \u7684\u80fd\u529b\u3002\u542c\u8d77\u6765\u4e0d\u505a\u767d\u4e0d\u505a\u3002\u8fd9\u4e00 Prompt \u540e\u6765\u88ab OpenAI \u62ff\u53bb\u6539\u4e86\u4e00\u6539\u53d8\u6210\u4e86\u300aLet’s verify step by step\u300b\uff0c\u800c\u8fd9\u7bc7\u8bba\u6587\u662f\u76ee\u524d\u6240\u6709\u60f3\u8981\u4e86\u89e3 o1 \u7684\u4eba\u53cd\u590d\u9605\u8bfb\u7684\u6838\u5fc3\u3002<\/p>\n

\u5f53\u7136\u4ec5\u4ec5\u4f9d\u9760\u63d0\u793a\u8bcd\u5de5\u7a0b\u6784\u5efa CoT \u4e0d\u53ef\u80fd\u662f o1 \u5982\u6b64\u5f3a\u5927\u7684\u539f\u56e0\uff0c\u4f46 CoT \u8fd9\u79cd\u4e00\u6b65\u6b65\u63a8\u8fdb\u903b\u8f91\u7684\u505a\u6cd5\uff0c\u6210\u4e3a\u5927\u6a21\u578b\u589e\u5f3a\u63a8\u7406\u80fd\u529b\u7684\u4e3b\u6d41\u65b9\u5411\u3002<\/p>\n

 <\/p>\n

03 CoT + Supervised Fine-Tune<\/h2>\n

 <\/p>\n

\u5f53\u7136\uff0c\u4e5f\u6709\u4eba\u5c1d\u8bd5\u5c06 CoT \u7684\u591a\u6b65\u63a8\u7406\u80fd\u529b\u4f7f\u7528 SFT \u7684\u65b9\u5f0f\u6559\u7ed9 LLM\u3002\u300aSTaR: Bootstrapping Reasoning With Reasoning\u300b\u662f\u4e00\u4e2a\u65e9\u671f\u5c1d\u8bd5\u3002\u4e0b\u56fe\u6765\u81ea\u8be5\u8bba\u6587\uff1a
\n<\/p>\n

\u8bba\u6587\u7684\u601d\u8def\u662f\u8fd9\u6837\u7684\u3002\u9996\u5148\u6211\u4eec\u7528\u4e0a\u6587\u7684\u63d0\u793a\u8bcd\u5de5\u7a0b\u7684\u65b9\u5f0f\uff0c\u8ba9\u6a21\u578b\u5c1d\u8bd5 CoT \u5728\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u63a8\u7406\uff0c\u4f1a\u5f97\u5230\u4e00\u6279\u7b54\u6848\uff0c\u7b54\u6848\u81ea\u7136\u6709\u5bf9\u6709\u9519\uff1a<\/p>\n

-\u5982\u679c\u5f97\u5230\u4e86\u6b63\u786e\u7684\u7b54\u6848\uff0c\u6211\u4eec\u8ba4\u4e3a\u6a21\u578b\u4ea7\u751f\u7684\u5bf9\u5e94\u7684 CoT \u662f\u4f18\u8d28 CoT\uff0c\u90a3\u4e48\u628a\u8fd9\u6837\u7684\u4f18\u8d28\u7684\u201c\u95ee\u9898-CoT-\u7b54\u6848\u201d\u6837\u672c\u6536\u96c6\u8d77\u6765\uff0c\u5c31\u5f97\u5230\u4e00\u4e2a\u65b0\u7684\u6570\u636e\u96c6\uff0c\u4f7f\u7528\u8be5\u6570\u636e\u96c6\u53bb SFT \u6211\u4eec\u7684 LLM\uff0c\u4e0d\u65ad\u5faa\u73af\uff0c\u5c31\u53ef\u4ee5\u5f97\u5230\u63a8\u7406\u80fd\u529b\u66f4\u5f3a\u7684 LLM\uff1b<\/p>\n

-\u800c\u5982\u679c\u6709\u4e00\u4e9b\u95ee\u9898\u4e0a\uff0cLLM \u59cb\u7ec8\u7b54\u9519\uff0c\u90a3\u6211\u4eec\u76f4\u63a5\u8ba9 LLM \u770b\u5230\u201c\u95ee\u9898+\u7b54\u6848\u201d\uff0c\u8ba9\u5b83\u751f\u6210\u4e00\u4e2a\u4ece\u95ee\u9898\u5230\u7b54\u6848\u7684 CoT\uff0c\u6211\u4eec\u53ef\u4ee5\u8ba4\u4e3a\uff0c\u5728\u5df2\u77e5\u7b54\u6848\u7684\u60c5\u51b5\u4e0b\uff0cLLM \u751f\u6210\u7684 CoT \u662f\u6b63\u786e\u7684\uff0c\u4e8e\u662f\u8fd9\u4e00\u90e8\u5206\u201c\u95ee\u9898-CoT-\u7b54\u6848\u201d\u7684\u6837\u672c\u4e5f\u53ef\u4ee5\u7528\u4e8e\u8bad\u7ec3\u3002<\/p>\n

\u7531\u4e8e\u8fd9\u7bc7\u7814\u7a76\u65f6\u95f4\u8f83\u65e9\uff0c\u6211\u4eec\u73b0\u5728\u5f88\u5bb9\u6613\u53d1\u73b0\u5176\u4e2d\u7684\u6f0f\u6d1e\uff0c\u6bd4\u5982\uff1aLLM \u5176\u5b9e\u7ecf\u5e38\u53d1\u751f\u201c\u8fc7\u7a0b\u9519\u8bef\u4f46\u662f\u7ed3\u679c\u6b63\u786e\u201d\u6216\u8005\u201c\u8fc7\u7a0b\u6b63\u786e\u4f46\u662f\u7ed3\u679c\u9519\u8bef\u201d\u7684\u60c5\u51b5\uff0c\u8fd9\u610f\u5473\u7740\u4e0a\u6587\u6211\u4eec\u7528\u4e8e\u8bad\u7ec3\u7684\u6837\u672c\u5176\u5b9e\u5e76\u6ca1\u6709\u90a3\u4e48\u9ad8\u7684\u8d28\u91cf\u3002\u90a3\u4e48\u5982\u4f55\u5f97\u5230\u66f4\u52a0\u6b63\u786e\u7684\u63a8\u7406\u8fc7\u7a0b\u5462\uff1f<\/p>\n

 <\/p>\n

04 Monte Carlo Tree Search<\/h2>\n

 <\/p>\n

\u4e0a\u6587\u6211\u4eec\u5df2\u7ecf\u77e5\u9053 CoT \u628a\u4ece\u95ee\u9898\u5230\u7b54\u6848\u7684\u903b\u8f91\u5df2\u7ecf\u62c6\u6210\u4e86\u4e00\u4e2a\u53c8\u4e00\u4e2a\u4e2d\u95f4\u601d\u7ef4\u8fc7\u7a0b\uff0c\u90a3\u4e48 MCTS \u662f\u5426\u53ef\u4ee5\u7528\u4e8e\u641c\u7d22\u4e0b\u4e00\u6b65\u63a8\u7406\u7684\u6700\u4f73\u601d\u7ef4\u6b65\u9aa4\uff0c\u4ece\u800c\u83b7\u5f97\u6700\u4f73\u63a8\u7406\u601d\u7ef4\u94fe\uff1f\u81ea\u7136\u53ef\u4ee5\u3002<\/p>\n

\u300aMutual Reasoning Makes Smaller LLMs Stronger Problem-Solvers\u300b\u5c31\u8bbe\u8ba1\u4e86\u8fd9\u6837\u4e00\u4e2a MCTS \u7b97\u6cd5\uff0c\u79f0\u4e3a rStar\uff0c\u4ed6\u4eec\u5c06\u9879\u76ee\u5f00\u6e90\u81f3\u4e86GitHub\u3002\u4e0b\u56fe\u6765\u81ea\u8be5\u8bba\u6587\uff0c\u662f\u4e0d\u662f\u548c\u4e0a\u6587 MCTS \u7684\u56fe\u7247\u6709\u4e00\u4e9b\u76f8\u4f3c\u4e4b\u5904\u4e86\uff1f<\/p>\n

<\/p>\n

\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u7814\u7a76\u4eba\u5458\u628a CoT \u7684\u4e2d\u95f4\u6b65\u9aa4\u5206\u6210\u4e86 5 \u4e2a\u7c7b\u578b\u7684\u7ed3\u70b9\uff1a<\/p>\n

1. \u4ea7\u751f\u4e0b\u4e00\u6b65\u63a8\u7406
\n2. \u4ea7\u751f\u540e\u7eed\u6240\u6709\u63a8\u7406
\n3. \u4ea7\u751f\u4e00\u4e2a\u5b50\u95ee\u9898\u53ca\u7b54\u6848
\n4. \u91cd\u65b0\u56de\u7b54\u5b50\u95ee\u9898
\n5. \u91cd\u6784\u95ee\u9898
\n\u7136\u540e\u5229\u7528 MCTS \u51b3\u5b9a\u4e0b\u4e00\u4e2a\u601d\u7ef4\u6b65\u9aa4\u7ed3\u70b9\u3002\u7531\u4e00\u4e2a\u53c8\u4e00\u4e2a\u601d\u7ef4\u7ed3\u70b9\u94fe\u63a5\u8d77\u6765\u7684\u8def\u5f84\uff0c\u5c31\u662f CoT\u3002\u6211\u4eec\u7b80\u5355\u62ff\u6240\u6709\u5f97\u5230\u7684\u6700\u7ec8\u7ed3\u679c\u8fdb\u884c\u6295\u7968\uff0c\u5c31\u80fd\u5f97\u5230\u4e00\u4e2a\u4e0d\u9519\u7684\u7ed3\u679c\u3002<\/p>\n

\u5f53\u7136\uff0c\u4f5c\u8005\u7814\u7a76\u4e0d\u6b62\u5982\u6b64\uff0c\u5982\u4e0a\u6587\u6240\u8bf4\uff0c\u5fc5\u987b\u8981\u80fd\u591f\u8861\u91cf\u6bcf\u4e00\u6b65\u7ed3\u70b9\u7684\u6b63\u786e\u6027\u548c\u63a8\u7406\u7684\u6b63\u786e\u6027\uff0c\u7814\u7a76\u4eba\u5458\u8bbe\u8ba1\u4e86\u5982\u4e0b\u65b9\u6cd5\uff1a<\/p>\n

-Discriminator \u7b5b\u9009\uff1a\u5728\u5f97\u5230\u7684\u539f\u59cb\u63a8\u7406\u8def\u5f84\u540e\uff0c\u968f\u673a Mask \u4e00\u90e8\u5206\uff0c\u7136\u540e\u4f7f\u7528\u53e6\u4e00\u4e2a\u6a21\u578b\u8fdb\u884c\u8f93\u51fa\uff0c\u5982\u679c\u5f97\u5230\u548c\u539f\u59cb Generator \u76f8\u540c\u7684\u7ed3\u679c\uff0c\u5219\u8bf4\u660e\u539f\u59cb\u63a8\u7406\u8def\u5f84\u53ef\u9760\u3002<\/p>\n

-\u7b54\u6848\u6b63\u786e\u6027\uff1a\u5c06\u6240\u6709\u7684\u6700\u7ec8\u7b54\u6848\u6536\u96c6\u8d77\u6765\uff0c\u67d0\u4e00\u4e2a\u7b54\u6848\u5360\u6240\u6709\u7b54\u6848\u7684\u6bd4\u4f8b\u5373\u4e3a\u7b54\u6848\u5f97\u5206\u3002<\/p>\n

-\u8fc7\u7a0b\u6b63\u786e\u6027\uff1a\u8def\u5f84\u4e2d\u6bcf\u4e00\u4e2a\u63a8\u7406\u7ed3\u70b9\uff0c\u5e76\u884c\u751f\u6210\u82e5\u5e72 2 \u578b\u7ed3\u70b9\u751f\u6210\u82e5\u5e72\u4e00\u6b65\u5230\u4f4d\u7684\u6700\u7ec8\u7ed3\u679c\uff0c\u8fd9\u4e9b\u7ed3\u679c\u4e2d\uff0c\u76ee\u524d\u8def\u5f84\u7684\u6700\u7ec8\u7ed3\u679c\u7684\u6bd4\u4f8b\u89c6\u4e3a\u8be5\u63a8\u7406\u7ed3\u70b9\u7684\u8fc7\u7a0b\u5f97\u5206\u3002\u901a\u8fc7\u4e09\u90e8\u5206\u8861\u91cf\uff0c\u5c31\u80fd\u5f97\u5230\u4e00\u6761\u6700\u4f73\u8def\u5f84\uff0c\u5c06\u6700\u4f73\u8def\u5f84\u7684\u6700\u7ec8\u7ed3\u679c\u89c6\u4e3a MCTS \u7684\u7ed3\u679c\u3002<\/p>\n

 <\/p>\n

05 Generator + Verifier<\/h2>\n

 <\/p>\n

\u9664\u4e86\u4e0a\u6587\u7684 MCTS \u53ef\u4ee5\u5c06\u601d\u7ef4\u8fc7\u7a0b\u7ec4\u7ec7\u6210\u6811\u5f62\u5e76\u8fdb\u884c\u63a2\u7d22\uff0c\u8fd8\u6709\u5176\u4ed6\u65b9\u5f0f\u3002\u6bd4\u5982\u5f3a\u5316\u5b66\u4e60\uff0c\u6211\u4eec\u518d\u6b21\u770b\u5f3a\u5316\u5b66\u4e60\u7684\u4ecb\u7ecd\uff1a
\n<\/p>\n

\u5982\u679c\u6211\u4eec\u628a LLM \u4f5c\u4e3a Actor\uff0c\u53e6\u4e00\u4e2a\u9488\u5bf9\u95ee\u9898\u8bad\u7ec3\u7684 RM \u4f5c\u4e3a Environment\uff0c\u52a0\u4e0a\u4e00\u4e2a\u9690\u542b\u7684 Critic\u3002\u8fd9\u6837\u4e00\u4e2a\u5f3a\u5316\u5b66\u4e60\u7684\u5faa\u73af\u5c31\u662f\uff1aActor \u9488\u5bf9\u95ee\u9898\u4ea7\u751f\u4e00\u4e2a\u7ed3\u679c\uff0cRM \u9a8c\u8bc1\u8fd9\u4e2a\u7ed3\u679c\u7684\u6b63\u786e\u6027\u5e76\u53cd\u9988\u7ed9 Agent\uff0cActor \u548c Critic \u6839\u636e Reward \u8fdb\u884c\u5b66\u4e60\u8bad\u7ec3\u3002\u6211\u4eec\u628a\u8fd9\u91cc\u7684 Agent \u79f0\u4e3a Generator\uff0c\u56e0\u4e3a\u4ed6\u7684\u4efb\u52a1\u662f\u4ea7\u751f\u7ed3\u679c\uff1bRM \u79f0\u4e3a Verifier\uff0c\u56e0\u4e3a\u4ed6\u7684\u4efb\u52a1\u662f\u9a8c\u8bc1\u7ed3\u679c\u3002<\/p>\n

\u4ed4\u7ec6\u60f3\u60f3\uff0cAgent \u5185\u90e8\u7684 Actor \u548c Critic \u7684\u5173\u7cfb\uff0c\u4e0e AlphaZero \u4f7f\u7528\u7684 Policy \u548c Value \u7f51\u7edc\u662f\u4e0d\u662f\u5f88\u50cf\uff1f\u4e8b\u5b9e\u4e5f\u662f\u8fd9\u6837\uff0cPolicy \u548c Value \u7f51\u7edc\u4e5f\u7b26\u5408 Actor \u548c Critic \u6846\u67b6\u3002<\/p>\n

\u73b0\u5728\u6211\u4eec\u603b\u7ed3\u4e00\u4e0b\uff0c\u4e00\u6b21\u5f3a\u5316\u5b66\u4e60\u7684\u8fc7\u7a0b\u6709 Actor\u3001Critic \u548c RM \u4e09\u4e2a\u7f51\u7edc\u53c2\u4e0e\u3002\u800c\u5728\u90e8\u7f72\u65f6\uff0c\u6839\u636e\u60c5\u51b5\u7684\u4e0d\u540c\uff0c\u4f7f\u7528\u4e0d\u540c\u7684\u6846\u67b6\u90e8\u7f72\uff1a\u68cb\u7c7b\u6e38\u620f\u4e2d\u53ea\u6709\u5728\u5bf9\u5c40\u7ed3\u675f\u7684\u65f6\u5019\uff0c\u624d\u80fd\u77e5\u9053\u80dc\u8d1f\uff0cRM \u7ed9\u4e88\u7684 Reward \u592a\u5c11\uff0c\u56e0\u6b64\u6211\u4eec\u9009\u62e9\u5728\u90e8\u7f72\u65f6\u4fdd\u7559 Actor-Critic \u6846\u67b6\uff0c\u7136\u540e\u8fdb\u884c MCTS \u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u89e3\u51b3\u65b9\u6848\uff1b\u800c\u5728 LLM \u90e8\u7f72\u65f6\uff0c\u6211\u4eec\u8bad\u7ec3\u597d\u7684 RM \u53ef\u4ee5\u53ca\u65f6\u53cd\u9988\uff0c\u90a3\u4e48\u6211\u4eec\u5728\u90e8\u7f72\u7684\u65f6\u5019\u81ea\u7136\u4e5f\u53ef\u4ee5\u7ec4\u5408 Actor \u548c RM \u6210 Generator-Verifier \u6846\u67b6\u3002<\/p>\n

OpenAI \u5728 GPT3\uff08ChatGPT \u662f\u57fa\u4e8e GPT-3.5 \u6a21\u578b\u7684\uff09\u65f6\u4ee3\u5c31\u5df2\u7ecf\u5728\u7814\u7a76\u8fd9\u4e2a\u65b9\u5411\u4e86\u3002\u4ed6\u4eec\u7ed9\u51fa\u7684\u65b9\u6848\u662f\u8bba\u6587\u300aTraining Verifiers to Solve Math Word Problems\u300b\u3002\u4e0b\u56fe\u6765\u81ea\u8be5\u8bba\u6587\uff1a<\/p>\n

<\/p>\n

\u4e0a\u56fe\u5bf9\u6bd4\u4e86\uff1a\u201c\u4ec5\u4ec5\u5fae\u8c03 Generator \u5f97\u5230\u7684\u7ed3\u679c\u7684\u6b63\u786e\u7387\u201d \u548c \u201c\u5fae\u8c03\u4e00\u4e2a Verifier\uff0c\u5bf9 Generator \u4ea7\u751f\u7684\u591a\u4e2a\u7ed3\u679c\u8fdb\u884c\u8bc4\u4f30\uff0c\u9009\u62e9\u66f4\u9ad8\u8bc4\u4ef7\u7684\u7ed3\u679c\u7684\u6b63\u786e\u7387\u201d\uff0c\u8bc1\u660e\u4e86 Verifier \u7684\u6709\u6548\u6027\u3002<\/p>\n

\u56e0\u4e3a\u8fd9\u91cc\u7684\u4efb\u52a1\u662f\uff1a\u6839\u636e\u95ee\u9898\u63a8\u7406\u5f97\u5230\u7ed3\u679c\u3002\u6240\u4ee5\u4f7f\u7528\u7684 Generator \u5e76\u4e0d\u4ea7\u751f\u4e2d\u95f4\u7684\u63a8\u7406\u8fc7\u7a0b\u800c\u662f\u76f4\u63a5\u4ea7\u751f\u7ed3\u679c\uff0cVerifier \u4e5f\u662f\u6211\u4eec\u5728\u5f3a\u5316\u5b66\u4e60\u4e00\u8282\u4e2d\u63d0\u5230\u7684 ORM\uff08\u57fa\u4e8e\u7ed3\u679c\u7684\u5956\u52b1\u6a21\u578b\uff09\uff0c\u5176\u4f5c\u7528\u662f\u6839\u636e Generator \u7684\u7ed3\u679c\u4ea7\u751f\u4e00\u4e2a\u8bc4\u5206\u3002\u6240\u4ee5\u8fd9\u91cc\u5e76\u6ca1\u6709\u6d89\u53ca\u6211\u4eec\u60f3\u8981\u63a2\u7a76\u7684\u591a\u6b65\u63a8\u7406\u7684\u8fc7\u7a0b\uff0c\u4ec5\u4ec5\u53ea\u662f\u53d1\u73b0 ORM \u9a8c\u8bc1\u5f97\u5230\u7684\u6700\u7ec8\u7ed3\u679c\u6bd4\u7b80\u5355\u5fae\u8c03\u66f4\u597d\u3002<\/p>\n

\u4e8e\u662f OpenAI \u56e2\u961f\u66f4\u8fdb\u4e00\u6b65\uff1a\u4e00\u65b9\u9762\u8ba9 Generator \u4e0d\u518d\u76f4\u63a5\u8f93\u51fa\u7ed3\u679c\uff0c\u800c\u662f\u4ea7\u751f\u9010\u6b65\u63a8\u7406\uff1b\u53e6\u4e00\u65b9\u9762\uff0c\u8bad\u7ec3\u4e86\u4e00\u4e2a PRM\uff08\u57fa\u4e8e\u8fc7\u7a0b\u7684\u5956\u52b1\u6a21\u578b\uff09\uff0c\u4f5c\u4e3a Verifier\uff0c\u5176\u4f5c\u7528\u662f\u9488\u5bf9 Generator \u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u6bcf\u4e00\u6b65\u4ea7\u751f\u8bc4\u5206\u3002\u6211\u4eec\u8ba4\u4e3a\u8fd9\u6837\u529b\u6c42 Generator \u63a8\u7406\u7684\u8fc7\u7a0b\u6b63\u786e\u6027\u6240\u4ea7\u751f\u7684\u7ed3\u679c\uff0c\u624d\u662f\u6700\u6709\u53ef\u80fd\u6b63\u786e\u7684\u3002<\/p>\n

\u8fd9\u5c31\u662f\u4e0a\u6587\u6211\u4eec\u63d0\u5230\u7684\u300aLet’s verify step by step\u300b\u3002\u5728\u8fd9\u4e2a\u5de5\u4f5c\u4e2d\uff0c\u56e2\u961f\u5bf9\u6bd4\u4e86 PRM \u548c ORM \u5206\u522b\u4f5c\u4e3a Verifier\uff0c\u641c\u7d22\u76f8\u540c Generator \u4ea7\u751f\u7684\u63a8\u7406\u7ed3\u679c\uff08\u6b64\u65f6\u4ed6\u4eec\u7684 Generator \u5df2\u7ecf\u662f GPT-4 \u4e86\uff09\uff0c\u8bc1\u660e PRM \u4f5c\u4e3a Verifier \u641c\u7d22\u5f97\u5230\u7684\u7ed3\u679c\u66f4\u51c6\u786e\u3002\u4e0b\u56fe\u6765\u81ea\u8be5\u8bba\u6587\uff1a<\/p>\n

<\/p>\n

\u4e0a\u56fe\u8bf4\u660e\uff0c\u540c\u4e00\u4e2a\u9010\u6b65\u63a8\u7406\u7684 Generator \u4ea7\u751f\u7684\u7ed3\u679c\uff0c\u6211\u4eec\u4f7f\u7528 ORM \u4f5c\u4e3a Verifier \u9009\u51fa\u7ed3\u679c\u6700\u4f73\u7684\u7b54\u6848\u662f\u6709\u6548\u7684\uff0c\u4f46\u6211\u4eec\u4f7f\u7528 PRM \u4f5c\u4e3a Verifier \u9009\u51fa\u8fc7\u7a0b\u6700\u4f73\u7684\u7b54\u6848\uff0c\u6b63\u786e\u7387\u66f4\u9ad8\uff01<\/p>\n

\u8fd9\u662f\u4e0d\u662f\u5c31\u662f\u6211\u4eec\u60f3\u8981\u5bfb\u627e\u7684 o1 \u80cc\u540e\u7684\u6280\u672f\u5462\uff1f\u6211\u4eec\u76ee\u524d\u53ea\u80fd\u731c\u6d4b\uff0c\u8fd9\u662f\u80cc\u540e\u7684\u6838\u5fc3\u6280\u672f\u4e4b\u4e00\u3002\u7406\u7531\u5982\u4e0b\uff1a<\/p>\n

1\u3001\u8fd9\u7bc7\u8bba\u6587\u8ddd\u79bb o1 \u7684\u53d1\u5e03\u65f6\u95f4\u6bd4\u8f83\u8fdc\uff0c\u4e00\u5e74\u7684\u65f6\u95f4\u8db3\u591f OpenAI \u7684\u7814\u7a76\u4eba\u5458\u66f4\u52a0\u6df1\u5165\u7814\u7a76\u8fd9\u4e2a\u65b9\u5411\u3002\u56e0\u4e3a PRM \u7684\u6709\u6548\u6027\uff0c\u867d\u7136\u4e00\u5e74\u65f6\u95f4\u4e5f\u8db3\u591f\u8c03\u6574\u81f3\u5176\u4ed6\u65b9\u5411\uff0c\u4f46\u6211\u4eec\u8fd8\u662f\u8ba4\u4e3a\u4ed6\u4eec\u662f\u6df1\u5165\u800c\u975e\u6389\u5934\u3002<\/p>\n

2\u3001\u6587\u7ae0\u8bc1\u660e\u4e86 PRM \u4f5c\u4e3a Verifier \u7684\u6709\u6548\u6027\uff0c\u663e\u7136\u53ef\u4ee5\u8fdb\u884c\u7684\u4e0b\u4e00\u6b65\u5c1d\u8bd5\u662f\uff1a\u5229\u7528\u5f3a\u5927\u7684 Verifier \u6539\u5584 Generator \u4ee5\u4ea7\u751f\u66f4\u4f18\u8d28\u7684\u7ed3\u679c\u3002\u4f46\u662f\u8bba\u6587\u5e76\u6ca1\u6709\u8fd9\u65b9\u9762\u63a8\u8fdb\uff0c\u6211\u4eec\u6709\u7406\u7531\u76f8\u4fe1 OpenAI \u5fc5\u5b9a\u505a\u4e86\u5c1d\u8bd5\uff0c\u4ea7\u751f\u7684\u7ed3\u679c\u662f\u5426\u5c31\u662f o1 \u4e0d\u5f97\u800c\u77e5\u3002<\/p>\n

\u8bf4\u5b8c\u731c\u6d4b\uff0c\u6211\u4eec\u7ee7\u7eed\u63a2\u7a76\u5229\u7528 Verifier \u8fdb\u884c\u641c\u7d22\u7684\u5176\u4ed6\u65b9\u5f0f\u3002\u6765\u81ea Google DeepMind \u4eca\u5e74 8 \u6708\u4efd\u7684\u6587\u7ae0\u300aScaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters\u300b\u505a\u4e86\u66f4\u591a\u7684\u7814\u7a76\u3002\u8fd9\u7bc7\u6587\u7ae0\u88ab\u8bb8\u591a\u4eba\u8ba4\u4e3a\u662f\u5c55\u793a\u4e86\u548c o1 \u80cc\u540e\u539f\u7406\u7c7b\u4f3c\u7684\u6280\u672f\u8def\u7ebf\u3002\u4e0b\u56fe\u6765\u81ea\u8be5\u8bba\u6587\uff1a<\/p>\n

<\/p>\n

\u5df2\u77e5\u6211\u4eec\u6709\u4e86 Generator \u548c Verifier\uff0c\u5982\u4f55\u8ba9\u4ed6\u4eec\u4e92\u76f8\u914d\u5408\u5f97\u5230\u6700\u4f18\u7ed3\u679c\u5462\uff1f\u4e0a\u6587\u63d0\u5230\u7684\u4e00\u4e2a\u65b9\u6cd5\u662f\uff0cGenerator \u5e76\u884c\u91c7\u6837\u5f97\u5230\u591a\u4e2a\u7ed3\u679c\uff0cVerifier \u8bc4\u4f30\u540e\u9009\u62e9\u6700\u9ad8\u5206\u3002\u8fd9\u5c31\u662f\u4e0a\u56fe\u91cc\u7684\u5de6\u4fa7\u7684 Parallel Sampling + Best-of-N \u7684\u505a\u6cd5\u3002\u4f46\u662f\u663e\u7136\u8fd8\u6709\u5176\u4ed6\u505a\u6cd5\uff1a<\/p>\n

-\u5728\u4ea7\u751f\u591a\u4e2a\u7ed3\u679c\u7684\u65f6\u5019\uff0c\u9664\u4e86\u5e76\u884c\u91c7\u6837\u591a\u4e2a\u7ed3\u679c\uff0c\u4e5f\u53ef\u4ee5\u8ba9 Generator \u4ea7\u751f\u4e00\u4e2a\u7ed3\u679c\u4e4b\u540e\uff0c\u81ea\u5df1\u5bf9\u7ed3\u679c\u8fdb\u884c\u68c0\u67e5\u548c\u6539\u6b63\uff0c\u5f97\u5230\u4e00\u4e2a\u7b54\u6848\u5e8f\u5217\uff0c\u4ed6\u4eec\u4e4b\u95f4\u4e0d\u518d\u662f\u5e73\u884c\u7684\u5173\u7cfb\u3002<\/p>\n

-\u7531 Verifier \u8fdb\u884c\u9009\u62e9\u7684\u65f6\u5019\uff0c\u9664\u4e86 Best-of-N\uff0c\u4e5f\u53ef\u4ee5\u6709\u5176\u4ed6\u529e\u6cd5\u3002\u5982\u6765\u81ea\u8be5\u8bba\u6587\u7684\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n

<\/p>\n

\u8bba\u6587\u53d1\u73b0\uff0c\u7b80\u5355\u7684\u95ee\u9898\uff0c\u6211\u4eec\u5e94\u8be5\u4f7f\u7528 Verifier \u9f13\u52b1 Generator \u8fdb\u884c\u81ea\u6211\u68c0\u67e5\u548c\u6539\u6b63\uff0c\u800c\u4e0d\u662f\u76f2\u76ee\u5e76\u884c\u641c\u7d22\u3002\u800c\u590d\u6742\u7684\u95ee\u9898\uff0cGenerator \u5e76\u884c\u5c1d\u8bd5\u4e0d\u540c\u7684\u65b9\u6848\u662f\u66f4\u597d\u7684\u9009\u62e9\u3002<\/p>\n

\u7c7b\u4f3c\u7684\u5de5\u4f5c\u662f\u300aA Comparative Study on Reasoning Patterns of OpenAI’s o1 Model\u300b\u3002\u8bba\u6587\u56e2\u961f\u5728 GitHub \u4e0a\u5f00\u6e90\u4e86\u4e00\u4e2a\u590d\u523b o1 \u7684\u9879\u76ee Open-o1\uff0c\u8fd9\u7bc7\u6587\u7ae0\u662f\u4ed6\u4eec\u5728 o1 \u53d1\u5e03\u4e4b\u540e\u7684\u4e00\u4e9b\u8c03\u7814\u6210\u679c\u3002\u4e0b\u56fe\u6765\u81ea\u8be5\u8bba\u6587\uff1a<\/p>\n

<\/p>\n

\u56e2\u961f\u4f7f\u7528 GPT-4o \u4f5c\u4e3a\u9aa8\u67b6\u6a21\u578b\uff0c\u7136\u540e\u4f7f\u7528\u56db\u79cd\u5e38\u89c1\u7684\u8ba9 LLM \u5b9e\u73b0 thinking-before-reasoning \u7684\u65b9\u6cd5\uff0c\u5bf9\u6bd4\u4e86\u4ed6\u4eec\u7684\u6548\u679c\u3002\u56e2\u961f\u53d1\u73b0\uff0c\u5728 HotpotQA \u4efb\u52a1\u4e0a\uff0cBest-of-N \u548c Step-wise BoN \u7684\u65b9\u5f0f\u90fd\u80fd\u591f\u660e\u663e\u63d0\u9ad8 LLM \u7684\u63a8\u7406\u80fd\u529b\uff0cBoN \u751a\u81f3\u4f7f\u5f97 GPT-4o \u8d85\u8fc7\u4e86 o1 \u6a21\u578b\u3002<\/p>\n

 <\/p>\n

06 OpenR<\/h2>\n

 <\/p>\n

\u5728\u5f53\u524d\u8bd5\u56fe\u590d\u523b o1 \u7684\u5f00\u6e90\u9879\u76ee\u4e2d\uff0cOpenR \u662f\u5176\u4e2d\u5b8c\u6210\u5ea6\u76f8\u5bf9\u8f83\u9ad8\u7684\u4e00\u4e2a\u3002
\n<\/p>\n

\u56fe\u7247\u6765\u81ea\u5176\u5b98\u65b9\u6587\u6863\uff0c\u76ee\u524d\u6765\u770b\uff0c\u5176\u5b9e\u73b0\u4e86\u6570\u636e\u6536\u96c6\u4ee5\u53ca\u7b26\u5408 Generator-Verifier \u6846\u67b6\u7684\u8bad\u7ec3\u4e0e\u90e8\u7f72\u3002<\/p>\n

\u6570\u636e\u6536\u96c6\u6839\u636e\u5b98\u65b9\u4ecb\u7ecd\uff0c\u6570\u636e\u6536\u96c6\u65b9\u6cd5\u6765\u81ea\u8bba\u6587\uff1a\u300aImprove Mathematical Reasoning in Language Models by Automated Process Supervision\u300b\u3002\u7b80\u5355\u6765\u8bf4\uff0c\u5c31\u662f\u4f7f\u7528 MCTS \u6269\u5c55\u539f\u6709\u7684 problem-final_answer \u6570\u636e\u96c6\uff0c\u751f\u6210 CoT \u63a8\u7406\u6b65\u9aa4\u3002\u6700\u540e\u5f97\u5230\u4e00\u4e2a MATH-APS \u6570\u636e\u96c6\u3002<\/p>\n

\u76f8\u5173\u6570\u636e\u96c6\u5df2\u6258\u7ba1\u5230ModelScope\u4e0a\uff1a<\/p>\n

PRM800K-Stepwise\u6570\u636e\u96c6\uff1a
\nhttps:\/\/modelscope.cn\/datasets\/AI-ModelScope\/openai-prm800k-stepwise-critic\/<\/p>\n

MATH-APS\u6570\u636e\u96c6\uff1a
\nhttps:\/\/modelscope.cn\/datasets\/AI-ModelScope\/MATH-APS\/<\/p>\n

Math-Shepherd\u6570\u636e\u96c6\uff1a
\nhttps:\/\/modelscope.cn\/datasets\/AI-ModelScope\/Math-Shepherd<\/p>\n

Generator \u8bad\u7ec3\u56e2\u961f\u4f7f\u7528\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684 PPO \u7b97\u6cd5\u53d8\u4f53\u8bad\u7ec3 Generator\u3002\u7b80\u5355\u6765\u8bf4\uff0cPPO \u7b97\u6cd5\u662f\u5229\u7528 Reward Model \u63d0\u4f9b\u7684 Reward \u4fe1\u606f\u8fdb\u884c\u8bad\u7ec3\uff0c\u540c\u65f6\u9650\u5236 Actor\uff0c\u4f7f\u5176\u5728\u5b66\u4e60\u8fc7\u7a0b\u4e2d\u4e0d\u504f\u79bb\u539f\u59cb Actor \u592a\u8fdc\uff0c\u907f\u514d\u4e22\u5931\u5df2\u6709\u7684\u77e5\u8bc6\u3002\u76ee\u524d OpenR \u652f\u6301 APPO\u3001GRPO \u548c TPPO \u4e09\u4e2a\u53d8\u4f53\u3002<\/p>\n

Virifier \u8bad\u7ec3\u56e2\u961f\u4f7f\u7528 SFT \u76d1\u7763\u5b66\u4e60\u8bad\u7ec3\u4e00\u4e2a PRM\uff0c\u4f7f\u7528\u7684\u6570\u636e\u96c6\u9664\u4e86\u4e0a\u6587\u7684 MATH-APS \u6570\u636e\u96c6\uff0c\u8fd8\u5305\u62ec PRM800K \u3001 Math-Shepherd \u4e24\u4e2a\u5f00\u6e90\u6570\u636e\u96c6\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5728\u8fd9\u4e09\u4e2a Step \u7ea7\u522b\u7684\u6570\u636e\u96c6\u4e0a\uff0c\u56e2\u961f\u5bf9\u6bcf\u4e00\u4e2a step \u6807\u6ce8 \u201c+\u201d \u6216\u8005 \u201c-\u201d \u6807\u7b7e\uff0c\u7136\u540e\u8ba9 PRM \u5b66\u4e60\u9884\u6d4b\u6bcf\u4e00\u4e2a Step \u7684\u6807\u7b7e\uff0c\u5224\u65ad\u5176\u6b63\u786e\u4e0e\u5426\u3002<\/p>\n

\u6a21\u578b\u4f7f\u7528”step-wise”\u6570\u636e\u8fdb\u884c PPO training\uff0c\u6240\u5f97\u5230\u7684\u6a21\u578b\u6743\u91cd\u5df2\u7ecf\u6258\u7ba1\u5230ModelScope\uff0c\u76ee\u524d\u63d0\u4f9b\u4e86 SFT\u3001PRM \u548c RL \u6a21\u578b\u7684checkpoint\uff0c\u4ee5\u53ca\u90e8\u5206GGUF\u683c\u5f0f\uff1a<\/p>\n

mistral-7b-sft\u6a21\u578b\uff1a
\nhttps:\/\/modelscope.cn\/models\/AI-ModelScope\/mistral-7b-sft<\/p>\n

RL\u6a21\u578b\uff08GGUF\u7248\uff09\uff1a
\nhttps:\/\/modelscope.cn\/models\/QuantFactory\/math-shepherd-mistral-7b-rl-GGUF<\/p>\n

PRM\u6a21\u578b\uff1a
\n-GGUF\u7248\uff1ahttps:\/\/modelscope.cn\/models\/QuantFactory\/math-shepherd-mistral-7b-prm-GGUF
\n-PRM\u6a21\u578b\uff1ahttps:\/\/modelscope.cn\/models\/AI-ModelScope\/math-shepherd-mistral-7b-prm<\/p>\n

\u63a8\u7406\u90e8\u7f72\u5728\u90e8\u7f72\u65f6\uff0cOpenR \u901a\u8fc7\u6307\u5b9a\u7684 Generator \u548c Verifier \u4f7f\u7528\u641c\u7d22\u7b97\u6cd5\uff0c\u83b7\u5f97\u63a8\u7406\u8fc7\u7a0b\u548c\u6700\u7ec8\u7b54\u6848\u3002\u76ee\u524d\u652f\u6301 MCTS\u3001Beam Search\u3001best_of_n \u4e09\u79cd\u641c\u7d22\u65b9\u5f0f\u3002
\n<\/p>\n

\u56fe\u7247\u6765\u81ea\u8bba\u6587\u300aOpenR: An Open Source Framework for Advanced Reasoning with Large Language Models\u300b\u3002OpenR \u7684\u7ed3\u6784\u5982\u56fe\u6240\u793a\uff0c\u76ee\u524d\u6765\u8bf4\uff0cOpenR \u5b9e\u73b0\u4e86\u4e00\u6761\u590d\u523b O1 \u7684\u94fe\u8def\uff0c\u4ece\u6536\u96c6\u8bad\u7ec3\u6570\u636e\u5230\u8bad\u7ec3\u4e00\u4e2a PRM\uff0c\u518d\u5230\u4f7f\u7528 PRM \u5f3a\u5316\u5b66\u4e60\uff0c\u6700\u540e\u90e8\u7f72\u6a21\u578b\u8fdb\u884c\u641c\u7d22\uff0c\u5e76\u4e14\u56e2\u961f\u5c06\u5de5\u4f5c\u5168\u90e8\u5f00\u6e90\u4f9b\u793e\u533a\u5b66\u4e60\u5c1d\u8bd5\uff0c\u8ba9\u6211\u4eec\u5f97\u4ee5\u4e00\u7aa5\u7a76\u7adf\u3002<\/p>\n

\u521b\u7a7a\u95f4\u4f53\u9a8c\u6211\u4eec\u5728\u9b54\u642d\u793e\u533a\u521b\u7a7a\u95f4\u90e8\u7f72\u4e86OpenR\u7684\u63a8\u7406\u670d\u52a1\uff0c\u5f00\u53d1\u8005\u53ef\u4ee5\u901a\u8fc7\u8bbf\u95ee\u4ee5\u4e0b\u94fe\u63a5\uff0c\u5728\u7ebf\u4f53\u9a8cOpenR\u7684\u6548\u679c\uff1ahttps:\/\/www.modelscope.cn\/studios\/modelscope\/OpenR_Inference<\/p>\n

 <\/p>\n

07 \u7ed3\u8bba<\/h2>\n

 <\/p>\n

\u6211\u4eec\u8c03\u7814\u7684\u4ee5\u4e0a\u6709\u5173\u591a\u6b65\u63a8\u7406\u7684\u8bba\u6587\uff0c\u8bc1\u660e\u4e86\u8ba9 LLM \u9010\u6b65\u8fdb\u884c\u63a8\u7406\u800c\u4e0d\u662f\u8df3\u8fc7\u4e2d\u95f4\u8fc7\u7a0b\u80fd\u591f\u663e\u8457\u589e\u5f3a\u5176\u5728\u548c\u903b\u8f91\u76f8\u5173\u7684\u95ee\u9898\u4e0a\u7684\u51c6\u786e\u6027\u3002\u4e3a\u4e86\u8ba9 LLM \u8fdb\u884c\u9010\u6b65\u63a8\u7406\uff0c\u6211\u4eec\u9664\u4e86\u53ef\u4ee5\u7528\u7b80\u5355\u7684\u63d0\u793a\u8bcd\u5de5\u7a0b\u5f15\u5bfc\uff0c\u4e5f\u53ef\u4ee5\u4f7f\u7528\u4e00\u4e9b\u5e26\u6709\u4e2d\u95f4\u8fc7\u7a0b\u7684\u6570\u636e\u96c6\u8fdb\u884c\u5fae\u8c03\u3002\u800c\u66f4\u6709\u6548\u7684\u505a\u6cd5\u662f\uff0c\u8bad\u7ec3\u4e00\u4e2a\u80fd\u591f\u9010\u6b65\u9a8c\u8bc1 Generator \u51c6\u786e\u6027\u7684 Verifier \u5bf9 Generator \u751f\u6210\u7684\u7ed3\u679c\u8fdb\u884c\u641c\u7d22\u3002<\/p>\n

\u4ece\u76ee\u524d\u7684\u731c\u6d4b\u548c\u8bba\u6587\u6765\u770b\uff0c\u8fc8\u5411 o1 \u7684\u53ef\u80fd\u6280\u672f\u6b63\u662f\u57fa\u4e8e\u5f3a\u5927\u7684 LLM Generator \u548c LLM Verifier \u4e4b\u95f4\u7684\u914d\u5408\u3002\u8fd9\u79cd\u5de6\u811a\u8e29\u53f3\u811a\uff0c\u81ea\u5df1\u5bf9\u6297\u81ea\u5df1\u8fdb\u884c\u81ea\u6211\u8fed\u4ee3\u7684\u505a\u6cd5\uff0c\u5728\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\u5df2\u7ecf\u4e0d\u662f\u7b2c\u4e00\u6b21\u4e86\uff0c\u4f46 OpenAI \u5374\u662f\u7387\u5148\u5c06\u8fd9\u6837\u7684\u6a21\u5f0f\u5f15\u5165\u5149\u8bad\u7ec3 Generator \u5c31\u8017\u8d44\u5de8\u5927\u7684 LLM \u9886\u57df\uff0c\u786e\u5b9e\u8d22\u5927\u6c14\u7c97\u3002<\/p>\n

\u56e0\u6b64\u6211\u4eec\u8ba4\u4e3a\uff0c\u5982\u679c\u60f3\u8981\u590d\u523b o1\uff0c\u9996\u5148\u9700\u8981\u7684\u662f\u4e00\u4e2a\u53ef\u4ee5\u4e3a Generator \u63d0\u4f9b\u8f85\u52a9\u548c\u6307\u5bfc\u7684 Verifier\uff0c\u800c\u4e3a\u4e86\u4ea7\u751f\u8bad\u7ec3 Verifier \u6240\u9700\u8981\u7684\u6570\u636e\uff0c\u53ef\u4ee5\u53c2\u8003\u4e0a\u6587 CoT + Supervised Fine-Tune \u548c Monte Carlo Tree Search \u7ae0\u8282\uff0c\u4ee5\u8f83\u4f4e\u6210\u672c\u7684\u65b9\u5f0f\u83b7\u5f97\u8f83\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u3002\u8fd9\u4e5f\u5c31\u662f\u4e3a\u4ec0\u4e48\u6211\u4eec\u4e5f\u4ecb\u7ecd\u4e86\u8fd9\u51e0\u9879\u5de5\u4f5c\u3002<\/p>\n

\u6700\u540e\u6211\u4eec\u4ecb\u7ecd\u4e86\u4e00\u4e2a\u5b8c\u6210\u5ea6\u8f83\u9ad8\u7684\u5f00\u6e90\u9879\u76ee\uff0c\u57fa\u4e8e\u4ed6\u4eec\u7684\u5de5\u4f5c\uff0c\u6211\u4eec\u5f97\u4ee5\u6574\u7406\u81ea\u5df1\u7684\u601d\u8def\u548c\u60f3\u6cd5\u3002<\/p>\n

 <\/p>\n

08 \u53c2\u8003<\/h2>\n

Chain of Thought Prompting Elicits Reasoning in Large Language Models
\nLarge Language Models are Zero-Shot Reasoners
\nSTaR: Bootstrapping Reasoning With Reasoning
\nMutual Reasoning Makes Smaller LLMs Stronger Problem-Solvers
\nTraining Verifiers to Solve Math Word Problems
\nLet’s verify step by step
\nScaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
\nA Comparative Study on Reasoning Patterns of OpenAI’s o1 Model
\nOpenR: An Open Source Framework for Advanced Reasoning with Large Language Models<\/p>\n","protected":false},"excerpt":{"rendered":"

2022 \u5e74 OpenAI \u53d1\u5e03\u4e86 ChatGPT\uff0c\u6210\u4e3a\u4e16\u754c\u4e0a\u6700\u5feb\u7a81\u7834\u4e0a\u4ebf\u7528\u6237\u7684 APP\uff0c\u90a3\u65f6\u5019\u4eba\u4eec\u90fd\u8ba4\u4e3a\uff0c\u6211\u4eec\u79bb\u771f\u6b63\u7684\u4eba\u5de5\u667a\u80fd\u66f4\u8fd1\u4e86\u3002\u4f46\u662f\u4eba\u4eec\u5f88\u5feb\u53d1\u73b0\uff0cChatGPT \u53ef\u4ee5\u5bf9\u8bdd\u804a\u5929\uff0c\u751a\u81f3\u53ef\u4ee5\u5199\u8bd7\u5199\u6587\u7ae0\uff0c\u4f46\u5728\u7b80\u5355\u7684\u903b\u8f91\u4e0a\u4f9d\u7136\u4e0d\u5c3d\u5982\u4eba\u610f\uff0c\u6bd4\u5982…<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[34],"tags":[],"class_list":["post-14235","post","type-post","status-publish","format-standard","hentry","category-knowledge"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/14235","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/comments?post=14235"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/14235\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=14235"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=14235"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=14235"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}