{"id":28377,"date":"2025-12-09T08:55:19","date_gmt":"2025-12-09T00:55:19","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=28377"},"modified":"2025-12-09T08:55:19","modified_gmt":"2025-12-09T00:55:19","slug":"deepseek-v3%e5%88%b0v3-2%e7%9a%84%e8%bf%9b%e5%8c%96%e4%b9%8b%e8%b7%af%ef%bc%8c%e4%b8%80%e6%96%87%e7%9c%8b%e5%85%a8","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2025\/12\/09\/28377\/","title":{"rendered":"DeepSeek V3\u5230V3.2\u7684\u8fdb\u5316\u4e4b\u8def\uff0c\u4e00\u6587\u770b\u5168"},"content":{"rendered":"
\u65e0\u8bba DeepSeek \u6a21\u578b\u4e0e\u5176\u4ed6\u8f83\u5c0f\u7684\u5f00\u653e\u6743\u91cd\u6a21\u578b\u6216\u50cf GPT-5.1 \u6216 Gemini 3.0 Pro \u8fd9\u6837\u7684\u4e13\u6709\u6a21\u578b\u76f8\u6bd4\u7684\u5e02\u573a\u4efd\u989d\u5982\u4f55\uff0c\u6709\u4e00\u4ef6\u4e8b\u662f\u80af\u5b9a\u7684\uff1aDeepSeek \u7684\u53d1\u5e03\u603b\u662f\u5f88\u6709\u8da3\uff0c\u800c\u4e14\u4ece\u968f\u5f00\u653e\u6743\u91cd\u6a21\u578b\u68c0\u67e5\u70b9\u4e00\u8d77\u53d1\u5e03\u7684\u6280\u672f\u62a5\u544a\u4e2d\u603b\u6709\u5f88\u591a\u503c\u5f97\u5b66\u4e60\u7684\u4e1c\u897f\u3002\u200b<\/div>\n
\n
\n

12 \u6708 1 \u65e5\uff0cDeepSeek \u4e00\u53e3\u6c14\u53d1\u5e03\u4e86\u4e24\u6b3e\u65b0\u6a21\u578b\uff1aDeepSeek-V3.2 \u548c DeepSeek-V3.2-Speciale\u3002<\/p>\n

\u51e0\u5929\u8fc7\u53bb\uff0c\u70ed\u5ea6\u4f9d\u65e7\u4e0d\u51cf\uff0c\u89e3\u8bfb\u5176\u6280\u672f\u62a5\u544a\u7684\u535a\u5ba2\u4e5f\u6b63\u5728\u4e0d\u65ad\u6d8c\u73b0\u3002\u77e5\u540d AI \u7814\u7a76\u8005\u548c\u535a\u4e3b Sebastian Raschka \u53d1\u5e03\u8fd9\u7bc7\u6df1\u5ea6\u535a\u5ba2\u5c24\u5176\u503c\u5f97\u4e00\u8bfb\uff0c\u5176\u8be6\u7ec6\u68b3\u7406\u4e86 DeepSeek V3 \u5230 V3.2 \u7684\u8fdb\u5316\u5386\u7a0b\u3002<\/p>\n

<\/p>\n

\u673a\u5668\u4e4b\u5fc3\u7f16\u8bd1\u4e86\u8fd9\u7bc7\u6df1\u5ea6\u6280\u672f\u535a\u5ba2\uff0c\u4ee5\u98e8\u8bfb\u8005\uff1a<\/p>\n

<\/p>\n

    \n
  • \u535a\u5ba2\u6807\u9898\uff1aA Technical Tour of the DeepSeek Models from V3 to V3.2<\/li>\n
  • \u535a\u5ba2\u5730\u5740\uff1ahttps:\/\/sebastianraschka.com\/blog\/2025\/technical-deepseek.html<\/li>\n<\/ul>\n

    \u4e0e DeepSeek V3 \u7684\u53d1\u5e03\u7b56\u7565\u5982\u51fa\u4e00\u8f99\uff0cDeepSeek \u56e2\u961f\u518d\u6b21\u9009\u62e9\u5728\u4e00\u4e2a\u7f8e\u56fd\u4e3b\u8981\u8282\u5047\u65e5\u5468\u672b\u53d1\u5e03\u4e86\u4ed6\u4eec\u7684\u65b0\u65d7\u8230\u6a21\u578b\u3002<\/p>\n

    \u9274\u4e8e DeepSeek V3.2 \u5c55\u73b0\u51fa\u4e86\u6781\u4f73\u7684\u6027\u80fd\u6c34\u5e73\uff08\u5728\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5bf9\u6807 GPT-5 \u548c Gemini 3.0 Pro\uff09\uff0c\u52a0\u4e4b\u5b83\u672c\u8eab\u4e5f\u662f\u4e00\u4e2a\u5f00\u653e\u6743\u91cd\u6a21\u578b\uff0c\u8fd9\u65e0\u7591\u503c\u5f97\u91cd\u70b9\u5173\u6ce8\u3002<\/p>\n

    <\/p>\n

    \u56fe 1\uff1aDeepSeek V3.2 \u4e0e\u4e13\u6709\u65d7\u8230\u6a21\u578b\u4e4b\u95f4\u7684\u57fa\u51c6\u6d4b\u8bd5\u5bf9\u6bd4\u3002\u6765\u81ea DeepSeek V3.2 \u62a5\u544a \u5e76\u52a0\u4e0a\u4e86\u6ce8\u91ca\u3002<\/p>\n

    \u6211\u66fe\u5728\u6587\u7ae0\u300a\u5927\u578b LLM \u67b6\u6784\u5bf9\u6bd4\u300b\u7684\u5f00\u7bc7\u4ecb\u7ecd\u8fc7\u5b83\u7684\u524d\u8eab \u2014\u2014DeepSeek V3\u3002\u968f\u7740\u65b0\u67b6\u6784\u7684\u4e0d\u65ad\u6d8c\u73b0\uff0c\u6211\u5728\u8fc7\u53bb\u51e0\u4e2a\u6708\u91cc\u4e00\u76f4\u5728\u6301\u7eed\u66f4\u65b0\u90a3\u7bc7\u6587\u7ae0\u3002<\/p>\n

    \u539f\u672c\uff0c\u6211\u521a\u548c\u5bb6\u4eba\u5ea6\u5b8c\u611f\u6069\u8282\u5047\u671f\u56de\u6765\uff0c\u8ba1\u5212\u300c\u53ea\u662f\u300d\u5728\u539f\u6587\u7ae0\u4e2d\u589e\u52a0\u4e00\u4e2a\u65b0\u7ae0\u8282\u6765\u7b80\u5355\u4ecb\u7ecd\u8fd9\u6b21\u65b0\u53d1\u5e03\u7684 DeepSeek V3.2\u3002\u4f46\u968f\u540e\u6211\u610f\u8bc6\u5230\uff0c\u8fd9\u6b21\u66f4\u65b0\u5305\u542b\u592a\u591a\u6709\u8da3\u7684\u4fe1\u606f\u548c\u7ec6\u8282\uff0c\u4ec5\u4ec5\u4e00\u4e2a\u7ae0\u8282\u65e0\u6cd5\u6db5\u76d6\u3002\u56e0\u6b64\uff0c\u6211\u51b3\u5b9a\u5c06\u5176\u5199\u6210\u4e00\u7bc7\u7bc7\u5e45\u8f83\u957f\u7684\u72ec\u7acb\u6587\u7ae0\u3002<\/p>\n

    \u4ed6\u4eec\u7684\u6280\u672f\u62a5\u544a\u4e2d\u6db5\u76d6\u4e86\u5927\u91cf\u6709\u8da3\u7684\u9886\u57df\u548c\u503c\u5f97\u5b66\u4e60\u7684\u77e5\u8bc6\uff0c\u8ba9\u6211\u4eec\u5f00\u59cb\u5427\uff01<\/p>\n

    \u76ee\u5f55<\/h4>\n

    1. DeepSeek \u53d1\u5e03\u65f6\u95f4\u7ebf<\/p>\n

    2. \u6df7\u5408\u63a8\u7406\u6a21\u578b\u4e0e\u4e13\u7528\u63a8\u7406\u6a21\u578b<\/p>\n

    3. \u4ece DeepSeek V3 \u5230 V3.1<\/p>\n

    \u00a0 3.1 DeepSeek V3 \u6982\u89c8\u4e0e\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b (MLA)<\/p>\n

    \u00a0 3.2 DeepSeek R1 \u6982\u89c8\u4e0e\u5e26\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60 (RLVR)<\/p>\n

    \u00a0 3.3 DeepSeek R1-0528 \u7248\u672c\u5347\u7ea7<\/p>\n

    \u00a0 3.4 DeepSeek V3.1 \u6df7\u5408\u63a8\u7406<\/p>\n

    4. DeepSeek V3.2-Exp \u4e0e\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236<\/p>\n

    5. \u91c7\u7528\u81ea\u6211\u9a8c\u8bc1\u4e0e\u81ea\u6211\u4fee\u6b63\u7684 DeepSeekMath V2<\/p>\n

    \u00a0 5.1 \u81ea\u6211\u9a8c\u8bc1<\/p>\n

    \u00a0 5.2 \u81ea\u6211\u4fee\u6b63<\/p>\n

    6. DeepSeek V3.2 (2025 \u5e74 12 \u6708 1 \u65e5)<\/p>\n

    \u00a0 6.1 DeepSeek V3.2 \u67b6\u6784<\/p>\n

    \u00a0 6.2 \u5f3a\u5316\u5b66\u4e60\u66f4\u65b0<\/p>\n

    \u00a0 6.3 GRPO \u66f4\u65b0<\/p>\n

    \u00a0 6.4 DeepSeek V3.2-Speciale \u4e0e\u6269\u5c55\u601d\u7ef4<\/p>\n

    7. \u603b\u7ed3<\/p>\n

    1. DeepSeek \u53d1\u5e03\u65f6\u95f4\u7ebf<\/h4>\n

    \u867d\u7136 DeepSeek V3 \u5728 2024 \u5e74 12 \u6708\u521a\u53d1\u5e03\u65f6\u5e76\u6ca1\u6709\u7acb\u523b\u5f15\u8d77\u8f70\u52a8\uff0c\u4f46\u968f\u540e\u63a8\u51fa\u7684 DeepSeek R1 \u63a8\u7406\u6a21\u578b\uff08\u57fa\u4e8e\u76f8\u540c\u7684\u67b6\u6784\uff0c\u4f7f\u7528 DeepSeek V3 \u4f5c\u4e3a\u57fa\u7840\u6a21\u578b\uff09\u6539\u53d8\u4e86\u5c40\u9762\u3002\u5b83\u5e2e\u52a9 DeepSeek \u6210\u4e3a\u4e86\u6700\u53d7\u6b22\u8fce\u7684\u5f00\u653e\u6743\u91cd\u6a21\u578b\u4e4b\u4e00\uff0c\u5e76\u6210\u4e3a\u4e86 OpenAI\u3001Google\u3001xAI \u548c Anthropic \u7b49\u516c\u53f8\u4e13\u6709\u6a21\u578b\u7684\u6709\u529b\u66ff\u4ee3\u65b9\u6848\u3002<\/p>\n

    <\/p>\n

    \u56fe 2\uff1a2024 \u5e74 12 \u6708\u53d1\u5e03\u7684 DeepSeek V3 \u548c R1 \u67b6\u6784\u3002\u6211\u4eec\u5c06\u5728\u7a0d\u540e\u7684\u7ae0\u8282\u4e2d\u91cd\u6e29\u5e76\u8ba8\u8bba\u8fd9\u4e9b\u67b6\u6784\u7ec6\u8282\u3002<\/p>\n

    \u90a3\u4e48\uff0c\u81ea V3\/R1 \u4ee5\u6765\u6709\u4ec0\u4e48\u65b0\u53d8\u5316\u5417\uff1f\u6211\u76f8\u4fe1 DeepSeek \u56e2\u961f\u4eca\u5e74\u4e00\u5b9a\u975e\u5e38\u5fd9\u788c\u3002\u7136\u800c\uff0c\u81ea DeepSeek R1 \u53d1\u5e03\u540e\u7684\u8fc7\u53bb 10 \u5230 11 \u4e2a\u6708\u91cc\uff0c\u5e76\u6ca1\u6709\u51fa\u73b0\u91cd\u5927\u7684\u7248\u672c\u53d1\u5e03\u3002<\/p>\n

    \u5c31\u6211\u4e2a\u4eba\u800c\u8a00\uff0c\u6211\u8ba4\u4e3a\u82b1\u5927\u7ea6\u4e00\u5e74\u7684\u65f6\u95f4\u6765\u53d1\u5e03\u4e00\u4e2a\u4e3b\u8981\u7684 LLM \u7248\u672c\u662f\u5408\u7406\u7684\uff0c\u56e0\u4e3a\u8fd9\u6d89\u53ca\u6d77\u91cf\u7684\u5de5\u4f5c\u3002\u7136\u800c\uff0c\u6211\u5728\u5404\u79cd\u793e\u4ea4\u5a92\u4f53\u5e73\u53f0\u4e0a\u770b\u5230\u4eba\u4eec\u5ba3\u79f0\u8be5\u56e2\u961f\u5df2\u7ecf\u300c\u51c9\u4e86\u300d\u3002<\/p>\n

    \u6700\u540e\uff0c\u8fd9\u5e76\u4e0d\u4ee3\u8868\u4ed6\u4eec\u4ec0\u4e48\u90fd\u6ca1\u53d1\u5e03\u3002\u4eca\u5e74\u5176\u5b9e\u9646\u7eed\u6709\u4e00\u4e9b\u8f83\u5c0f\u7684\u53d1\u5e03\uff0c\u4f8b\u5982 DeepSeek V3.1 \u548c V3.2-Exp\u3002<\/p>\n

    <\/p>\n

    \u56fe 3\uff1a\u81ea\u53bb\u5e74\u4ee5\u6765\u7684 DeepSeek \u53d1\u5e03\u60c5\u51b5\u3002\u4e3b\u8981\u6a21\u578b\u4ee5\u7ea2\u8272\u663e\u793a\u3002<\/p>\n

    \u6b63\u5982\u6211\u5728\u4e5d\u6708\u4efd\u6240\u9884\u6d4b\u7684\u90a3\u6837\uff0cDeepSeek V3.2-Exp \u7684\u53d1\u5e03\u65e8\u5728\u4e3a\u6258\u7ba1\u521a\u521a\u53d1\u5e03\u7684 V3.2 \u6a21\u578b\u51c6\u5907\u751f\u6001\u7cfb\u7edf\u548c\u63a8\u7406\u57fa\u7840\u8bbe\u65bd\u3002<\/p>\n

    V3.2-Exp \u548c V3.2 \u4f7f\u7528\u4e86\u4e00\u79cd\u975e\u6807\u51c6\u7684\u7a00\u758f\u6ce8\u610f\u529b\uff08Sparse Attention\uff09\u53d8\u4f53\uff0c\u8fd9\u9700\u8981\u5b9a\u5236\u4ee3\u7801\uff0c\u6211\u4eec\u7a0d\u540e\u4f1a\u8be6\u7ec6\u4ecb\u7ecd\u8fd9\u79cd\u673a\u5236\u3002<\/p>\n

    2. \u6df7\u5408\u63a8\u7406\u6a21\u578b\u4e0e\u4e13\u7528\u63a8\u7406\u6a21\u578b<\/h4>\n

    \u5728\u8fdb\u4e00\u6b65\u8ba8\u8bba\u6a21\u578b\u7ec6\u8282\u4e4b\u524d\uff0c\u6709\u5fc5\u8981\u5148\u63a2\u8ba8\u4e00\u4e0b\u6574\u4f53\u7684\u6a21\u578b\u7c7b\u578b\u3002\u6700\u521d\uff0cDeepSeek V3 \u662f\u4f5c\u4e3a\u4e00\u4e2a\u57fa\u7840\u6a21\u578b\u53d1\u5e03\u7684\uff0c\u800c DeepSeek R1 \u5219\u589e\u52a0\u4e86\u989d\u5916\u7684\u540e\u8bad\u7ec3\uff0c\u4ee5\u5f00\u53d1\u6210\u4e00\u4e2a\u4e13\u7528\u7684\u63a8\u7406\u6a21\u578b\u3002\u8fd9\u4e00\u8fc7\u7a0b\u603b\u7ed3\u5728\u4e0b\u56fe\u4e2d\u3002<\/p>\n

    <\/p>\n

    \u56fe 4\uff1aDeepSeek R1 \u8bad\u7ec3\u6d41\u7a0b\u6982\u89c8<\/p>\n

    \u8fd9\u91cc\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cDeepSeek V3 \u662f\u57fa\u7840\u6a21\u578b\uff0c\u800c DeepSeek R1 \u662f\u4e13\u7528\u7684\u63a8\u7406\u6a21\u578b\u3002<\/p>\n

    \u5728 DeepSeek \u63a8\u8fdb\u7684\u540c\u65f6\uff0c\u5176\u4ed6\u56e2\u961f\u4eca\u5e74\u4e5f\u53d1\u5e03\u4e86\u8bb8\u591a\u975e\u5e38\u5f3a\u5927\u7684\u5f00\u653e\u6743\u91cd\u63a8\u7406\u6a21\u578b\u3002\u4eca\u5e74\u6700\u5f3a\u7684\u5f00\u653e\u6743\u91cd\u6a21\u578b\u4e4b\u4e00\u662f Qwen3\u3002\u6700\u521d\uff0c\u5b83\u662f\u4f5c\u4e3a\u4e00\u4e2a\u6df7\u5408\u63a8\u7406\u6a21\u578b\u53d1\u5e03\u7684\uff0c\u8fd9\u610f\u5473\u7740\u7528\u6237\u53ef\u4ee5\u5728\u540c\u4e00\u4e2a\u6a21\u578b\u4e2d\u5207\u6362\u63a8\u7406\u6a21\u5f0f\u548c\u975e\u63a8\u7406\u6a21\u5f0f\u3002\uff08\u5728 Qwen3 \u7684\u6848\u4f8b\u4e2d\uff0c\u8fd9\u79cd\u5207\u6362\u662f\u901a\u8fc7\u5206\u8bcd\u5668\u6dfb\u52a0 \/ \u7701\u7565\u6807\u7b7e\u6765\u5b9e\u73b0\u7684\u3002\uff09<\/p>\n

    \u4ece\u90a3\u65f6\u8d77\uff0cLLM \u56e2\u961f\u53d1\u5e03\u4e86\u4e13\u7528\u63a8\u7406\u6a21\u578b\u548c\u6307\u4ee4 \/ \u63a8\u7406\u6df7\u5408\u6a21\u578b\uff08\u6709\u4e9b\u56e2\u961f\u751a\u81f3\u5728\u8fd9\u4e24\u8005\u4e4b\u95f4\u53cd\u590d\u6a2a\u8df3\uff09\uff0c\u5982\u4e0b\u56fe\u7684\u65f6\u95f4\u7ebf\u6240\u793a\u3002<\/p>\n

    <\/p>\n

    \u56fe 5\uff1a\u4eca\u5e74\u53d1\u5e03\u7684\u90e8\u5206\u63a8\u7406\u6a21\u578b\u548c\u6df7\u5408\u6a21\u578b\u7684\u65f6\u95f4\u7ebf\u3002<\/p>\n

    \u4f8b\u5982\uff0cQwen3 \u6700\u521d\u662f\u6df7\u5408\u6a21\u578b\u3002\u4f46\u968f\u540e\uff0cQwen \u56e2\u961f\u5206\u522b\u53d1\u5e03\u4e86\u72ec\u7acb\u7684\u6307\u4ee4\uff08Instruct\uff09\u6a21\u578b\u548c\u63a8\u7406\uff08Reasoning\uff09\u6a21\u578b\uff0c\u56e0\u4e3a\u8fd9\u4e24\u79cd\u6a21\u578b\u5f00\u53d1\u8d77\u6765\u66f4\u5bb9\u6613\uff0c\u4e14\u5728\u5404\u81ea\u7684\u7528\u4f8b\u4e2d\u6027\u80fd\u8868\u73b0\u66f4\u597d\u3002<\/p>\n

    \u6709\u4e9b\u6a21\u578b\uff08\u5982 OpenAI \u7684 gpt-oss\uff09\u4ec5\u63d0\u4f9b\u6df7\u5408\u53d8\u4f53\uff0c\u7528\u6237\u53ef\u4ee5\u901a\u8fc7\u7cfb\u7edf\u63d0\u793a\u8bcd\uff08System Prompt\uff09\u9009\u62e9\u63a8\u7406\u529b\u5ea6\uff08\u6211\u6000\u7591 GPT-5 \u548c GPT-5.1 \u4e5f\u662f\u7c7b\u4f3c\u7684\u5904\u7406\u65b9\u5f0f\uff09\u3002<\/p>\n

    \u800c\u5728 DeepSeek \u7684\u6848\u4f8b\u4e2d\uff0c\u4ed6\u4eec\u4f3c\u4e4e\u53cd\u5176\u9053\u800c\u884c\u4e4b\uff0c\u4ece\u4e13\u7528\u63a8\u7406\u6a21\u578b\uff08R1\uff09\u8f6c\u5411\u4e86\u6df7\u5408\u6a21\u578b\uff08V3.1 \u548c V3.2\uff09\u3002\u4e0d\u8fc7\uff0c\u6211\u6000\u7591 R1 \u4e3b\u8981\u662f\u4f5c\u4e3a\u4e00\u4e2a\u7814\u7a76\u9879\u76ee\uff0c\u65e8\u5728\u5f00\u53d1\u63a8\u7406\u65b9\u6cd5\u548c\u9a8c\u8bc1\u5f53\u65f6\u6700\u597d\u7684\u63a8\u7406\u6a21\u578b\u3002V3.2 \u7684\u53d1\u5e03\u53ef\u80fd\u66f4\u591a\u662f\u4e3a\u4e86\u5f00\u53d1\u9488\u5bf9\u4e0d\u540c\u7528\u4f8b\u7684\u6700\u4f73\u6574\u4f53\u6a21\u578b\u3002\uff08\u5728\u8fd9\u91cc\uff0cR1 \u66f4\u50cf\u662f\u4e00\u4e2a\u6d4b\u8bd5\u5e73\u53f0\u6216\u539f\u578b\u6a21\u578b\u3002\uff09<\/p>\n

    \u6211\u4e5f\u731c\u60f3\uff0c\u867d\u7136 DeepSeek \u56e2\u961f\u5f00\u53d1\u4e86\u5177\u6709\u63a8\u7406\u80fd\u529b\u7684 V3.1 \u548c V3.2\uff0c\u4f46\u4ed6\u4eec\u53ef\u80fd\u4ecd\u5728\u5f00\u53d1\u4e13\u95e8\u7684 R2 \u6a21\u578b\u3002<\/p>\n

    3. \u4ece DeepSeek V3 \u5230 V3.1<\/h4>\n

    \u5728\u66f4\u8be6\u7ec6\u5730\u8ba8\u8bba\u65b0\u7684 DeepSeek V3.2 \u53d1\u5e03\u4e4b\u524d\uff0c\u6211\u8ba4\u4e3a\u5148\u6982\u8ff0\u4e00\u4e0b\u4ece V3 \u5230 V3.1 \u7684\u4e3b\u8981\u53d8\u5316\u4f1a\u5f88\u6709\u5e2e\u52a9\u3002<\/p>\n

    3.1 DeepSeek V3 \u6982\u89c8\u4e0e\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b (MLA)<\/strong><\/p>\n

    \u6211\u5df2\u7ecf\u5728\u5176\u4ed6\u51e0\u7bc7\u6587\u7ae0\u4e2d\u975e\u5e38\u8be6\u7ec6\u5730\u8ba8\u8bba\u4e86 DeepSeek V3 \u548c R1\u3002<\/p>\n

    \u603b\u7ed3\u4e00\u4e0b\u8981\u70b9\uff0cDeepSeek V3 \u662f\u4e00\u4e2a\u57fa\u7840\u6a21\u578b\uff0c\u5b83\u4f7f\u7528\u4e86\u4e24\u4e2a\u503c\u5f97\u6ce8\u610f\u7684\u67b6\u6784\u7279\u6027\uff1a\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08MoE\uff09\u548c\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08MLA\uff09\u3002<\/p>\n

    \u6211\u60f3\u4f60\u6b64\u65f6\u53ef\u80fd\u5df2\u7ecf\u5bf9 MoE \u975e\u5e38\u719f\u6089\u4e86\uff0c\u6240\u4ee5\u6211\u8fd9\u91cc\u8df3\u8fc7\u4ecb\u7ecd\u3002<\/p>\n

    \u53e6\u4e00\u4e2a\u503c\u5f97\u6ce8\u610f\u7684\u4eae\u70b9\u662f MLA \u7684\u4f7f\u7528\u3002MLA \u5df2\u88ab\u7528\u4e8e DeepSeek V2\u3001V3 \u548c R1\uff0c\u5b83\u63d0\u4f9b\u4e86\u4e00\u79cd\u8282\u7701\u5185\u5b58\u7684\u7b56\u7565\uff0c\u7279\u522b\u9002\u5408\u4e0e KV \u7f13\u5b58\u642d\u914d\u4f7f\u7528\u3002MLA \u7684\u6838\u5fc3\u601d\u60f3\u662f\u5728\u5c06\u952e\uff08Key\uff09\u548c\u503c\uff08Value\uff09\u5f20\u91cf\u5b58\u50a8\u5230 KV \u7f13\u5b58\u4e4b\u524d\uff0c\u5148\u5c06\u5b83\u4eec\u538b\u7f29\u5230\u4e00\u4e2a\u4f4e\u7ef4\u7a7a\u95f4\u4e2d\u3002<\/p>\n

    \u5728\u63a8\u7406\u65f6\uff0c\u8fd9\u4e9b\u538b\u7f29\u7684\u5f20\u91cf\u5728\u4f7f\u7528\u524d\u4f1a\u88ab\u6295\u5f71\u56de\u5176\u539f\u59cb\u5927\u5c0f\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u8fd9\u867d\u7136\u589e\u52a0\u4e86\u4e00\u6b21\u989d\u5916\u7684\u77e9\u9635\u4e58\u6cd5\uff0c\u4f46\u663e\u8457\u51cf\u5c11\u4e86\u5185\u5b58\u4f7f\u7528\u3002<\/p>\n

    \uff08\u987a\u4fbf\u63d0\u4e00\u4e0b\uff0c\u67e5\u8be2 Query \u4e5f\u4f1a\u88ab\u538b\u7f29\uff0c\u4f46\u4ec5\u5728\u8bad\u7ec3\u671f\u95f4\uff0c\u63a8\u7406\u671f\u95f4\u4e0d\u4f1a\u3002\uff09<\/p>\n

    <\/p>\n

    \u56fe 6\uff1aDeepSeek V3 \u548c R1 \u4e2d\u7684\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b (MLA)\u3002\u4e3a\u7b80\u5355\u8d77\u89c1\uff0c\u672a\u663e\u793a\u67e5\u8be2\u5411\u91cf\u7684\u538b\u7f29\u7a7a\u95f4\u3002<\/p>\n

    \u4e0a\u56fe\u9610\u8ff0\u4e86 MLA \u80cc\u540e\u7684\u4e3b\u8981\u601d\u60f3\uff1a\u952e\u548c\u503c\u9996\u5148\u88ab\u6295\u5f71\u5230\u4e00\u4e2a\u6f5c\u5728\u5411\u91cf\u4e2d\uff0c\u8be5\u5411\u91cf\u53ef\u4ee5\u5b58\u50a8\u5728 KV \u7f13\u5b58\u4e2d\u4ee5\u51cf\u5c11\u5185\u5b58\u9700\u6c42\u3002\u8fd9\u9700\u8981\u7a0d\u540e\u8fdb\u884c\u5411\u4e0a\u6295\u5f71\uff08Up-projection\uff09\u56de\u5230\u539f\u59cb\u7684\u952e – \u503c\u7a7a\u95f4\uff0c\u4f46\u603b\u4f53\u4e0a\u5b83\u63d0\u9ad8\u4e86\u6548\u7387\uff08\u7c7b\u6bd4\u4e00\u4e0b\uff0c\u4f60\u53ef\u4ee5\u5c06\u5176\u60f3\u8c61\u4e3a\u00a0LoRA\u00a0\u4e2d\u7684\u964d\u7ef4\u548c\u5347\u7ef4\u6295\u5f71\uff09\u3002<\/p>\n

    \u987a\u4fbf\u8bf4\u4e00\u53e5\uff0c\u6b63\u5982\u524d\u9762\u63d0\u5230\u7684\uff0cMLA \u5728 DeepSeek V3 \u4e2d\u5e76\u4e0d\u662f\u65b0\u4e8b\u7269\uff0c\u56e0\u4e3a\u5b83\u7684\u524d\u8eab DeepSeek V2 \u4e5f\u4f7f\u7528\u4e86\uff08\u751a\u81f3\u5f15\u5165\u4e86\uff09\u5b83\u3002<\/p>\n

    3.2 DeepSeek R1 \u6982\u89c8\u4e0e\u5e26\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60 (RLVR)<\/strong><\/p>\n

    DeepSeek R1 \u4f7f\u7528\u4e86\u4e0e\u4e0a\u8ff0 DeepSeek V3 \u76f8\u540c\u7684\u67b6\u6784\u3002\u533a\u522b\u5728\u4e8e\u8bad\u7ec3\u914d\u65b9\u3002\u5373\uff0c\u4f7f\u7528 DeepSeek V3 \u4f5c\u4e3a\u57fa\u7840\u6a21\u578b\uff0cDeepSeek R1 \u4e13\u6ce8\u4e8e\u300c\u5e26\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60\u300d\uff08RLVR\uff09\u65b9\u6cd5\uff0c\u4ee5\u63d0\u9ad8\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n

    RLVR \u7684\u6838\u5fc3\u601d\u60f3\u662f\u8ba9\u6a21\u578b\u4ece\u53ef\u4ee5\u8fdb\u884c\u7b26\u53f7\u5316\u6216\u7f16\u7a0b\u9a8c\u8bc1\u7684\u54cd\u5e94\u4e2d\u5b66\u4e60\uff0c\u4f8b\u5982\u6570\u5b66\u548c\u4ee3\u7801\uff08\u4f46\u8fd9\u5f53\u7136\u4e5f\u53ef\u4ee5\u6269\u5c55\u5230\u8fd9\u4e24\u4e2a\u9886\u57df\u4e4b\u5916\uff09\u3002<\/p>\n

    <\/p>\n

    \u56fe 7\uff1a\u4e00\u4e2a\u53ef\u9a8c\u8bc1\u4efb\u52a1\u7684\u793a\u4f8b\u3002<\/p>\n

    GRPO\u00a0\u7b97\u6cd5\uff0c\u5168\u79f0\u300c\u7fa4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\u300d\uff08Group Relative Policy Optimization\uff09\uff0c\u672c\u8d28\u4e0a\u662f\u300c\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\u300d\uff08PPO\uff09\u7b97\u6cd5\u7684\u4e00\u4e2a\u7b80\u5316\u53d8\u4f53\u3002PPO \u5728\u7528\u4e8e LLM \u5bf9\u9f50\u7684\u300c\u5e26\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\u300d\uff08RLHF\uff09\u4e2d\u975e\u5e38\u6d41\u884c\u3002<\/p>\n

    <\/p>\n

    \u56fe 8\uff1aLLM \u8bad\u7ec3\u4e2d\u5f3a\u5316\u5b66\u4e60\u8bbe\u7f6e\u7684\u6bd4\u8f83\u3002\u4f20\u7edf\u7684\u5e26 PPO \u7684 RLHF \u4f7f\u7528\u5956\u52b1\u6a21\u578b\uff08\u57fa\u4e8e\u4eba\u7c7b\u504f\u597d\u8bad\u7ec3\uff09\u548c\u8bc4\u8bba\u5bb6\uff08\u4ef7\u503c\u6a21\u578b\uff09\u6765\u6307\u5bfc\u5b66\u4e60\u3002GRPO \u53d6\u6d88\u4e86\u8bc4\u8bba\u5bb6\u6a21\u578b\u3002\u5e26 GRPO \u7684 RLVR \u66f4\u8fdb\u4e00\u6b65\uff0c\u79fb\u9664\u4e86\u5956\u52b1\u6a21\u578b\uff0c\u8f6c\u800c\u4f9d\u8d56\u6765\u81ea\u7b26\u53f7\u5de5\u5177\uff08\u5982\u8ba1\u7b97\u5668\u6216\u7f16\u8bd1\u5668\uff09\u7684\u53ef\u9a8c\u8bc1\u5956\u52b1\u3002<\/p>\n

    3.3 DeepSeek R1-0528 \u7248\u672c\u5347\u7ea7<\/strong><\/p>\n

    \u6b63\u5982 DeepSeek \u56e2\u961f\u81ea\u5df1\u6240\u8ff0\uff0cDeepSeek R1-0528 \u57fa\u672c\u4e0a\u662f\u4e00\u4e2a\u300c\u5c0f\u7248\u672c\u5347\u7ea7\u300d\u3002<\/p>\n

    \u67b6\u6784\u4e0e DeepSeek V3\/R1 \u4fdd\u6301\u4e00\u81f4\uff0c\u6539\u8fdb\u4e3b\u8981\u5728\u8bad\u7ec3\u65b9\u9762\uff0c\u4ee5\u4f7f\u5176\u8fbe\u5230\u5f53\u65f6 OpenAI o3 \u548c Gemini 2.5 Pro \u7684\u6c34\u5e73\u3002<\/p>\n

    \u9057\u61be\u7684\u662f\uff0cDeepSeek \u56e2\u961f\u6ca1\u6709\u53d1\u5e03\u4efb\u4f55\u5177\u4f53\u4fe1\u606f\u6765\u63cf\u8ff0\u8fd9\u662f\u5982\u4f55\u5b9e\u73b0\u7684\uff1b\u7136\u800c\uff0c\u4ed6\u4eec\u8868\u793a\u8fd9\u90e8\u5206\u6e90\u4e8e\u540e\u8bad\u7ec3\u6d41\u7a0b\u7684\u4f18\u5316\u3002\u6b64\u5916\uff0c\u6839\u636e\u5df2\u5206\u4eab\u7684\u4fe1\u606f\uff0c\u6211\u8ba4\u4e3a\u8be5\u6a21\u578b\u7684\u6258\u7ba1\u7248\u672c\u5728\u63a8\u7406\u65f6\u53ef\u80fd\u4f1a\u4f7f\u7528\u66f4\u591a\u7684\u8ba1\u7b97\u8d44\u6e90\uff08\u5373\u8fdb\u884c\u66f4\u957f\u65f6\u95f4\u7684\u63a8\u7406\uff09\u3002<\/p>\n

    3.4 DeepSeek V3.1 \u6df7\u5408\u63a8\u7406<\/strong><\/p>\n

    DeepSeek V3.1 \u662f\u4e00\u4e2a\u517c\u5177\u901a\u7528\u804a\u5929\uff08\u6307\u4ee4\uff09\u548c\u63a8\u7406\u80fd\u529b\u7684\u6df7\u5408\u6a21\u578b\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u4e0d\u518d\u5f00\u53d1\u4e24\u4e2a\u72ec\u7acb\u7684\u6a21\u578b\uff0c\u800c\u662f\u73b0\u5728\u6709\u4e00\u4e2a\u6a21\u578b\uff0c\u7528\u6237\u53ef\u4ee5\u901a\u8fc7\u804a\u5929\u63d0\u793a\u6a21\u677f\u5207\u6362\u6a21\u5f0f\uff08\u7c7b\u4f3c\u4e8e\u6700\u521d\u7684 Qwen3 \u6a21\u578b\uff09\u3002<\/p>\n

    DeepSeek V3.1 \u57fa\u4e8e DeepSeek V3.1-Base\uff0c\u800c\u540e\u8005\u53c8\u57fa\u4e8e DeepSeek V3\u3002\u5b83\u4eec\u90fd\u5171\u4eab\u76f8\u540c\u7684\u67b6\u6784\u3002<\/p>\n

    4. DeepSeek V3.2-Exp \u4e0e\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236<\/h4>\n

    DeepSeek V3.2-Exp (2025 \u5e74 9 \u6708) \u5f00\u59cb\u53d8\u5f97\u66f4\u6709\u8da3\u4e86\u3002<\/p>\n

    \u6700\u521d\uff0cDeepSeek V3.2-Exp \u5728\u53d1\u5e03\u65f6\u5e76\u6ca1\u6709\u9738\u699c\u57fa\u51c6\u6d4b\u8bd5\uff0c\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u5f53\u65f6\u56f4\u7ed5\u8fd9\u4e2a\u6a21\u578b\u7684\u5174\u594b\u5ea6\u4e0d\u9ad8\u7684\u539f\u56e0\u3002\u7136\u800c\uff0c\u6b63\u5982\u6211\u5728\u4e5d\u6708\u4efd\u63a8\u6d4b\u7684\u90a3\u6837\uff0c\u8fd9\u5f88\u53ef\u80fd\u662f\u4e00\u4e2a\u65e9\u671f\u7684\u5b9e\u9a8c\u6027\u53d1\u5e03\uff0c\u65e8\u5728\u4e3a\u66f4\u5927\u89c4\u6a21\u7684\u53d1\u5e03\u51c6\u5907\u57fa\u7840\u8bbe\u65bd\uff08\u7279\u522b\u662f\u63a8\u7406\u548c\u90e8\u7f72\u5de5\u5177\uff09\uff0c\u56e0\u4e3a DeepSeek V3.2-Exp \u4e2d\u6709\u4e00\u4e9b\u67b6\u6784\u4e0a\u7684\u53d8\u5316\u3002\u66f4\u5927\u7684\u53d1\u5e03\u662f DeepSeek V3.2\uff08\u4e0d\u662f V4\uff09\uff0c\u8fd9\u4e00\u90e8\u5206\u7a0d\u540e\u4f1a\u8be6\u7ec6\u4ecb\u7ecd\u3002<\/p>\n

    \u90a3\u4e48\uff0cDeepSeek V3.2-Exp \u6709\u4ec0\u4e48\u65b0\u4e1c\u897f\uff1f\u9996\u5148\uff0cDeepSeek V3.2-Exp \u662f\u57fa\u4e8e DeepSeek V3.1-Terminus \u4f5c\u4e3a\u57fa\u7840\u6a21\u578b\u8bad\u7ec3\u7684\u3002\u4ec0\u4e48\u662f DeepSeek V3.1-Terminus\uff1f\u5b83\u53ea\u662f\u4e0a\u4e00\u8282\u63d0\u5230\u7684 DeepSeek V3.1 \u68c0\u67e5\u70b9\u7684\u4e00\u4e2a\u5c0f\u5e45\u6539\u8fdb\u7248\u3002<\/p>\n

    \u6280\u672f\u62a5\u544a\u6307\u51fa\uff1a<\/p>\n

    \n

    DeepSeek-V3.2-Exp\uff0c\u4e00\u4e2a\u5b9e\u9a8c\u6027\u7684\u7a00\u758f\u6ce8\u610f\u529b\u6a21\u578b\uff0c\u5b83\u901a\u8fc7\u6301\u7eed\u8bad\u7ec3\u4e3a DeepSeek-V3.1-Terminus \u914d\u5907\u4e86 DeepSeek \u7a00\u758f\u6ce8\u610f\u529b (DSA)\u3002\u51ed\u501f\u7531 Lightning Indexer \u9a71\u52a8\u7684\u7ec6\u7c92\u5ea6\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236 DSA\uff0cDeepSeek-V3.2-Exp \u5728\u8bad\u7ec3\u548c\u63a8\u7406\u65b9\u9762\u90fd\u5b9e\u73b0\u4e86\u663e\u8457\u7684\u6548\u7387\u63d0\u5347\uff0c\u7279\u522b\u662f\u5728\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u4e2d\u3002<\/p>\n<\/blockquote>\n

    \u5982\u4e0a\u6bb5\u6240\u8ff0\uff0c\u8fd9\u91cc\u7684\u4e3b\u8981\u521b\u65b0\u662f\u4ed6\u4eec\u5728\u5bf9 DeepSeek V3.1-Terminus \u8fdb\u884c\u8fdb\u4e00\u6b65\u8bad\u7ec3\u4e4b\u524d\uff0c\u6dfb\u52a0\u4e86\u00a0DeepSeek \u7a00\u758f\u6ce8\u610f\u529b (DSA)\u00a0\u673a\u5236\u3002<\/p>\n

    \u8fd9\u4e2a DSA \u7531 (1) \u4e00\u4e2a lightning indexer \u548c (2) \u4e00\u4e2a Token \u9009\u62e9\u5668\uff08token-selector\uff09\u7ec4\u6210\uff0c\u76ee\u6807\u662f\u6709\u9009\u62e9\u5730\u51cf\u5c11\u4e0a\u4e0b\u6587\u4ee5\u63d0\u9ad8\u6548\u7387\u3002<\/p>\n

    \u4e3a\u4e86\u89e3\u91ca\u5b83\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8ba9\u6211\u4eec\u4ece\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u5f00\u59cb\u3002\u4f8b\u5982\uff0c\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u8fd9\u79cd\u6280\u672f\uff08\u6700\u8fd1\u88ab Gemma 3 \u548c Olmo 3 \u4f7f\u7528\uff09\u4f1a\u5c06\u6ce8\u610f\u529b\u7a97\u53e3\u9650\u5236\u4e3a\u56fa\u5b9a\u5927\u5c0f\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n

    <\/p>\n

    \u56fe 9\uff1a\u5728\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u4e2d\uff0c\u5f53\u524d\u7684\u67e5\u8be2 Token \u4e0d\u5173\u6ce8\u6240\u6709\u4e4b\u524d\u7684 Token\uff0c\u800c\u53ea\u5173\u6ce8\u4e00\u4e2a\u5b50\u96c6\u3002<\/p>\n

    DSA \u57fa\u4e8e\u4e0e\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u76f8\u540c\u7684\u60f3\u6cd5\uff1a\u53ea\u80fd\u5173\u6ce8\u4e00\u90e8\u5206\u8fc7\u53bb\u7684 Token\u3002\u7136\u800c\uff0cDSA \u4e0d\u662f\u901a\u8fc7\u56fa\u5b9a\u5bbd\u5ea6\u7684\u6ed1\u52a8\u7a97\u53e3\u6765\u9009\u62e9\u53ef\u5173\u6ce8\u7684 Token\uff0c\u800c\u662f\u62e5\u6709\u4e00\u4e2a\u7d22\u5f15\u5668\u548c Token \u9009\u62e9\u5668\u6765\u51b3\u5b9a\u54ea\u4e9b\u8fc7\u53bb\u7684 Token \u53ef\u4ee5\u88ab\u5173\u6ce8\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u53ef\u88ab\u5173\u6ce8\u7684 Token \u9009\u62e9\u66f4\u52a0\u968f\u673a\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n

    <\/p>\n

    \u56fe 10\uff1a\u5728 DSA \u4e2d\uff0c\u5f53\u524d Token \u53ef\u4ee5\u5173\u6ce8\u4e00\u7ec4\u9009\u5b9a\u7684\u8fc7\u53bb Token\uff0c\u800c\u4e0d\u662f\u50cf\u5e38\u89c4\u56e0\u679c\u6ce8\u610f\u529b\u90a3\u6837\u5173\u6ce8\u6240\u6709\u8fc7\u53bb Token\u3002<\/p>\n

    \u7136\u800c\uff0c\u867d\u7136\u6211\u4e0a\u9762\u8bf4\u4e86\u300c\u968f\u673a\u300d\uff0c\u4f46\u9009\u62e9\u54ea\u4e9b\u8fc7\u53bb Token \u7684\u6a21\u5f0f\u5b9e\u9645\u4e0a\u5e76\u4e0d\u662f\u968f\u673a\u7684\uff0c\u800c\u662f\u5b66\u4e60\u5f97\u5230\u7684\u3002<\/p>\n

    \u5b9e\u9645\u4e0a\uff0cDSA \u4f7f\u7528\u5176\u6240\u8c13\u7684 lightning indexer \u57fa\u4e8e\u6240\u6709\u5148\u524d\u7684 Token \u4e3a\u6bcf\u4e2a\u65b0\u7684\u67e5\u8be2 Token \u8ba1\u7b97\u76f8\u5173\u6027\u5206\u6570\u3002\u5bf9\u4e8e\u6b64\u8ba1\u7b97\uff0clightning indexer \u4f7f\u7528 DeepSeek \u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b (MLA) \u4e2d\u7684\u538b\u7f29 Token \u8868\u793a\uff0c\u5e76\u8ba1\u7b97\u4e0e\u5176\u4ed6 Token \u7684\u76f8\u4f3c\u5ea6\u3002\u76f8\u4f3c\u5ea6\u5206\u6570\u57fa\u672c\u4e0a\u662f\u901a\u8fc7 ReLU \u51fd\u6570\u7684\u67e5\u8be2\u5411\u91cf\u548c\u952e\u5411\u91cf\u4e4b\u95f4\u7684\u7f29\u653e\u70b9\u79ef\u3002<\/p>\n

    \u5982\u679c\u4f60\u5bf9\u6570\u5b66\u7ec6\u8282\u611f\u5174\u8da3\uff0c\u4e0b\u9762\u5c55\u793a\u4e86\uff08\u6458\u81ea\u8bba\u6587\u7684\uff09\u8be5 lightning indexer \u76f8\u4f3c\u5ea6\u5206\u6570\u7684\u65b9\u7a0b\uff1a<\/p>\n

    <\/p>\n

    \u8fd9\u91cc\uff0cw \u662f\u4e00\u4e2a\u5b66\u4e60\u5f97\u5230\u7684\u6bcf\u5934\uff08per-head\uff09\u52a0\u6743\u7cfb\u6570\uff0c\u51b3\u5b9a\u6bcf\u4e2a\u7d22\u5f15\u5668\u5934\u5bf9\u6700\u7ec8\u76f8\u4f3c\u5ea6\u5206\u6570\u7684\u8d21\u732e\u7a0b\u5ea6\u3002q \u6307\u7684\u662f\u67e5\u8be2\u5411\u91cf\uff0ck \u6307\u7684\u662f\u952e\u5411\u91cf\u3002\u4e0b\u9762\u662f\u4e0d\u540c\u4e0b\u6807\u7684\u5217\u8868\uff1a<\/p>\n

      \n
    • t\uff1a\u5f53\u524d\u67e5\u8be2 Token \u7684\u4f4d\u7f6e\uff1b<\/li>\n
    • s\uff1a\u5e8f\u5217\u4e2d\u5148\u524d Token \u7684\u4f4d\u7f6e (0 \u2264 s < t)\uff1b<\/li>\n
    • j\uff1a\u4e0d\u540c\u7d22\u5f15\u5668\u5934\u7684\u7d22\u5f15\uff08\u4e3a\u4e86\u7b80\u5355\u8d77\u89c1\uff0c\u56fe 10 \u4ec5\u663e\u793a\u4e86\u4e00\u4e2a\u5934\uff09\uff0c\u6240\u4ee5 q_{t,j} \u610f\u601d\u662f\u300c\u7d22\u5f15\u5668\u5934 j \u4e2d\u5f53\u524d Token t \u7684\u67e5\u8be2\u5411\u91cf\u300d\u3002<\/li>\n<\/ul>\n

      \u4f60\u53ef\u80fd\u4f1a\u6ce8\u610f\u5230\u7d22\u5f15\u5668\u4ec5\u9488\u5bf9\u67e5\u8be2\uff0c\u800c\u4e0d\u9488\u5bf9\u952e\u3002\u8fd9\u662f\u56e0\u4e3a\u6a21\u578b\u53ea\u9700\u8981\u51b3\u5b9a\u6bcf\u4e2a\u65b0\u67e5\u8be2\u5e94\u8be5\u8003\u8651\u54ea\u4e9b\u8fc7\u53bb\u7684 Token\u3002\u952e\u5df2\u7ecf\u88ab\u538b\u7f29\u5e76\u5b58\u50a8\u5728 KV \u7f13\u5b58\u4e2d\uff0c\u56e0\u6b64\u7d22\u5f15\u5668\u4e0d\u9700\u8981\u518d\u6b21\u5bf9\u5b83\u4eec\u8fdb\u884c\u8bc4\u5206\u6216\u538b\u7f29\u3002<\/p>\n

      \u8fd9\u91cc\u7684 ReLU \u51fd\u6570\uff0c\u56e0\u4e3a\u5b83\u662f f(x) = max(x, 0)\uff0c\u4f1a\u5c06\u8d1f\u7684\u70b9\u79ef\u4f4d\u7f6e\u5f52\u96f6\uff0c\u8fd9\u7406\u8bba\u4e0a\u53ef\u4ee5\u5b9e\u73b0\u7a00\u758f\u6027\u3002\u4f46\u7531\u4e8e\u5b58\u5728\u5bf9\u4e0d\u540c\u5934\u7684\u6c42\u548c\uff0c\u7d22\u5f15\u5668\u5206\u6570\u5b9e\u9645\u4e0a\u4e3a 0 \u7684\u53ef\u80fd\u6027\u5f88\u5c0f\u3002\u7a00\u758f\u6027\u4e3b\u8981\u6765\u81ea\u4e8e\u5355\u72ec\u7684 Token \u9009\u62e9\u5668\u3002<\/p>\n

      \u5355\u72ec\u7684 Token \u9009\u62e9\u5668\u4ec5\u4fdd\u7559\u5c11\u91cf\u9ad8\u5206 Token\uff08\u4f8b\u5982\uff0ctop-k \u4f4d\u7f6e\uff09\uff0c\u5e76\u6784\u5efa\u4e00\u4e2a\u7a00\u758f\u6ce8\u610f\u529b\u63a9\u7801\uff0c\u63a9\u76d6\u6389\u672a\u5305\u542b\u5728\u9009\u5b9a\u5b50\u96c6\u4e2d\u7684\u5176\u4ed6 Token\u3002\uff08\u6ce8\u610f\u8fd9\u91cc\u7684 k \u662f top-k \u4e2d\u7684 k\uff0c\u4e0d\u8981\u4e0e\u4e0a\u9762\u65b9\u7a0b\u4e2d\u7528\u4e8e\u952e\u7684 k \u6df7\u6dc6\uff0c\u8fd9\u662f\u4e00\u4e2a\u8d85\u53c2\u6570\uff0c\u5728 DeepSeek \u56e2\u961f\u5206\u4eab\u7684\u6a21\u578b\u4ee3\u7801\u4e2d\u8bbe\u7f6e\u4e3a 2048\u3002\uff09<\/p>\n

      \u4e0b\u56fe\u4ee5\u6d41\u7a0b\u56fe\u7684\u5f62\u5f0f\u8bf4\u660e\u4e86\u6574\u4e2a\u8fc7\u7a0b\u3002<\/p>\n

      <\/p>\n

      \u56fe 11\uff1aDeepSeek V3.2 \u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\u7684\u53ef\u89c6\u5316\u603b\u7ed3\u3002<\/p>\n

      \u603b\u7ed3\u4e00\u4e0b\uff0c\u7d22\u5f15\u5668\u548c Token \u9009\u62e9\u5668\u7684\u7ed3\u679c\u662f\uff0c\u6bcf\u4e2a Token \u53ea\u5173\u6ce8\u6a21\u578b\u4e60\u5f97\u7684\u8ba4\u4e3a\u6700\u76f8\u5173\u7684\u51e0\u4e2a\u8fc7\u53bb Token\uff0c\u800c\u4e0d\u662f\u6240\u6709 Token \u6216\u56fa\u5b9a\u7684\u5c40\u90e8\u7a97\u53e3\u3002<\/p>\n

      \u8fd9\u91cc\u7684\u76ee\u6807\u4e0d\u662f\u4e3a\u4e86\u8d85\u8d8a DeepSeek V3.1-Terminus \u7684\u6027\u80fd\uff0c\u800c\u662f\u5728\u51cf\u5c11\u6027\u80fd\u8870\u51cf\uff08\u7531\u4e8e\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\uff09\u7684\u540c\u65f6\uff0c\u53d7\u76ca\u4e8e\u6548\u7387\u7684\u63d0\u5347\u3002<\/p>\n

      \u603b\u7684\u6765\u8bf4\uff0cDSA \u5c06\u6ce8\u610f\u529b\u673a\u5236\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u4ece\u4e8c\u6b21\u7684 O(L\u00b2) \u964d\u4f4e\u5230\u4e86\u7ebf\u6027\u7684 O(Lk)\uff0c\u5176\u4e2d L \u662f\u5e8f\u5217\u957f\u5ea6\uff0ck \u662f\u9009\u5b9a Token \u7684\u6570\u91cf\u3002<\/p>\n

      5. \u91c7\u7528\u81ea\u6211\u9a8c\u8bc1\u4e0e\u81ea\u6211\u4fee\u6b63\u7684 DeepSeekMath V2<\/h4>\n

      \u8ba8\u8bba\u5b8c DeepSeek V3.2-Exp\uff0c\u6211\u4eec\u8d8a\u6765\u8d8a\u63a5\u8fd1\u672c\u6587\u7684\u4e3b\u9898\uff1aDeepSeek V3.2\u3002\u7136\u800c\uff0c\u8fd8\u6709\u4e00\u4e2a\u62fc\u56fe\u9700\u8981\u5148\u8ba8\u8bba\u3002<\/p>\n

      2025 \u5e74 11 \u6708 27 \u65e5\uff08\u7f8e\u56fd\u7684\u611f\u6069\u8282\uff09\uff0c\u5c31\u5728 DeepSeek V3.2 \u53d1\u5e03\u524d 4 \u5929\uff0cDeepSeek \u56e2\u961f\u53d1\u5e03\u4e86\u57fa\u4e8e\u00a0DeepSeek V3.2-Exp-Base\u00a0\u7684\u00a0DeepSeekMath V2\u3002<\/p>\n

      \u8be5\u6a21\u578b\u662f\u4e13\u95e8\u4e3a\u6570\u5b66\u5f00\u53d1\u7684\uff0c\u5e76\u5728\u591a\u4e2a\u6570\u5b66\u7ade\u8d5b\u4e2d\u83b7\u5f97\u4e86\u91d1\u724c\u7ea7\u7684\u5206\u6570\u3002\u672c\u8d28\u4e0a\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06\u5176\u89c6\u4e3a DeepSeek V3.2 \u7684\u6982\u5ff5\u9a8c\u8bc1\u6a21\u578b\uff0c\u5b83\u5f15\u5165\u4e86\u53e6\u4e00\u9879\u6280\u672f\u3002<\/p>\n

      \u8fd9\u91cc\u7684\u5173\u952e\u5728\u4e8e\uff0c\u63a8\u7406\u6a21\u578b\uff08\u5982 DeepSeek R1 \u548c\u5176\u4ed6\u6a21\u578b\uff09\u662f\u4f7f\u7528\u5916\u90e8\u9a8c\u8bc1\u5668\u8bad\u7ec3\u7684\uff0c\u6a21\u578b\u5b66\u4f1a\u4e86\u5728\u5f97\u51fa\u6700\u7ec8\u7b54\u6848\u4e4b\u524d\u81ea\u884c\u7f16\u5199\u89e3\u91ca\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u89e3\u91ca\u53ef\u80fd\u662f\u4e0d\u6b63\u786e\u7684\u3002<\/p>\n

      \u6b63\u5982 DeepSeek \u56e2\u961f\u7b80\u6d01\u6307\u51fa\u7684\u90a3\u6837\uff0c\u5e38\u89c4 RLVR \u7684\u7f3a\u70b9\u662f\uff1a<\/p>\n

      \n

      [\u2026] \u6b63\u786e\u7684\u7b54\u6848\u5e76\u4e0d\u4fdd\u8bc1\u6b63\u786e\u7684\u63a8\u7406\u3002<\/p>\n

      [\u2026] \u6a21\u578b\u53ef\u80fd\u901a\u8fc7\u6709\u7f3a\u9677\u7684\u903b\u8f91\u6216\u5e78\u8fd0\u7684\u9519\u8bef\u5f97\u51fa\u6b63\u786e\u7684\u7b54\u6848\u3002<\/p>\n<\/blockquote>\n

      \u4ed6\u4eec\u65e8\u5728\u89e3\u51b3\u7684 DeepSeek R1 RLVR \u65b9\u6cd5\u7684\u53e6\u4e00\u4e2a\u5c40\u9650\u6027\u662f\uff1a<\/p>\n

      \n

      [\u2026] \u8bb8\u591a\u6570\u5b66\u4efb\u52a1\uff08\u5982\u5b9a\u7406\u8bc1\u660e\uff09\u9700\u8981\u4e25\u683c\u7684\u9010\u6b65\u63a8\u5bfc\u800c\u4e0d\u662f\u6570\u503c\u7b54\u6848\uff0c\u8fd9\u4f7f\u5f97\u6700\u7ec8\u7b54\u6848\u5956\u52b1\u5e76\u4e0d\u9002\u7528\u3002<\/p>\n<\/blockquote>\n

      \u56e0\u6b64\uff0c\u4e3a\u4e86\u6539\u5584\u4e0a\u8ff0\u4e24\u4e2a\u7f3a\u70b9\uff0c\u4ed6\u4eec\u8bad\u7ec3\u4e86\u4e24\u4e2a\u6a21\u578b\uff1a<\/p>\n

        \n
      1. \u4e00\u4e2a\u7528\u4e8e\u5b9a\u7406\u8bc1\u660e\u7684\u57fa\u4e8e LLM \u7684\u9a8c\u8bc1\u5668\uff08Verifier\uff09\u3002<\/li>\n
      2. \u4e3b\u6a21\u578b\uff0c\u4e00\u4e2a\u8bc1\u660e\u751f\u6210\u5668\uff08Proof-Generator\uff09\uff0c\u5b83\u4f7f\u7528\u57fa\u4e8e LLM \u7684\u9a8c\u8bc1\u5668\u4f5c\u4e3a\u5956\u52b1\u6a21\u578b\uff08\u800c\u4e0d\u662f\u7b26\u53f7\u9a8c\u8bc1\u5668\uff09\u3002<\/li>\n<\/ol>\n

        \u9664\u4e86\u4e0a\u8ff0\u901a\u8fc7 LLM \u8fdb\u884c\u81ea\u6211\u9a8c\u8bc1\u5916\uff0c\u4ed6\u4eec\u8fd8\u4f7f\u7528\u4e86\u81ea\u6211\u4fee\u6b63\uff0c\u8ba9 LLM \u8fed\u4ee3\u5730\u6539\u8fdb\u5176\u81ea\u5df1\u7684\u7b54\u6848\u3002<\/p>\n

        5.1 \u81ea\u6211\u9a8c\u8bc1<\/strong><\/p>\n

        \u62e5\u6709\u4e00\u4e2a\u5bf9\u4e2d\u95f4\u6b65\u9aa4\u8fdb\u884c\u8bc4\u5206\u7684 LLM \u5e76\u4e0d\u662f\u4ec0\u4e48\u65b0\u9c9c\u4e8b\u3002\u6709\u4e0d\u5c11\u5173\u4e8e\u6240\u8c13\u300c\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\u300d\uff08Process Reward Models\uff09\u7684\u7814\u7a76\u4e13\u6ce8\u4e8e\u6b64\u3002\u4f8b\u5b50\u5305\u62ec\uff1aSolving Math Word Problems With Process- and Outcome-based Feedback (2022) \u6216 Let\u2019s Verify Step by Step (2023)\uff0c\u8fd8\u6709\u66f4\u591a\u3002<\/p>\n

        \u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\u7684\u6311\u6218\u5728\u4e8e\uff0c\u68c0\u67e5\u4e2d\u95f4\u5956\u52b1\u662f\u5426\u6b63\u786e\u5e76\u4e0d\u5bb9\u6613\uff0c\u8fd9\u4e5f\u53ef\u80fd\u5bfc\u81f4\u5956\u52b1\u9ed1\u5ba2\u653b\u51fb\uff08reward hacking\uff09\u3002<\/p>\n

        \u5728 2025 \u5e74 1 \u6708\u7684 DeepSeek R1 \u8bba\u6587\u4e2d\uff0c\u4ed6\u4eec\u6ca1\u6709\u4f7f\u7528\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff0c\u56e0\u4e3a\u4ed6\u4eec\u53d1\u73b0\uff1a<\/p>\n

        \n

        \u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\uff0c\u4e0e\u5176\u5728\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\u8fc7\u7a0b\u4e2d\u5f15\u5165\u7684\u989d\u5916\u8ba1\u7b97\u5f00\u9500\u76f8\u6bd4\uff0c\u5b83\u7684\u4f18\u52bf\u662f\u6709\u9650\u7684\u3002<\/p>\n<\/blockquote>\n

        \u5728\u6b64\u6587\u4e2d\uff0c\u4ed6\u4eec\u4ee5\u81ea\u6211\u9a8c\u8bc1\u7684\u5f62\u5f0f\u6210\u529f\u5730\u91cd\u65b0\u5ba1\u89c6\u4e86\u8fd9\u4e00\u70b9\u3002\u5176\u52a8\u673a\u662f\uff0c\u5373\u4f7f\u4e0d\u5b58\u5728\u53c2\u8003\u7b54\u6848\uff0c\u4eba\u7c7b\u5728\u9605\u8bfb\u8bc1\u660e\u548c\u53d1\u73b0\u95ee\u9898\u65f6\u4e5f\u80fd\u8fdb\u884c\u81ea\u6211\u7ea0\u6b63\u3002<\/p>\n

        \u56e0\u6b64\uff0c\u4e3a\u4e86\u5f00\u53d1\u4e00\u4e2a\u66f4\u597d\u7684\u64b0\u5199\u6570\u5b66\u8bc1\u660e\u7684\u6a21\u578b\uff08\u4e0b\u56fe\u4e2d\u7684 LLM 1\uff09\uff0c\u4ed6\u4eec\u5f00\u53d1\u4e86\u4e00\u4e2a\u8bc1\u660e\u9a8c\u8bc1\u5668\uff08\u4e0b\u56fe\u4e2d\u7684 LLM 2\uff09\uff0c\u5b83\u53ef\u4ee5\u7528\u4f5c\u300cLLM \u5373\u88c1\u5224\u300d\uff08LLM-as-a-judge\uff09\u6765\u5bf9\u8bc1\u660e\u8005\uff08LLM 1\uff09\u7684\u8f93\u51fa\u8fdb\u884c\u8bc4\u5206\u3002<\/p>\n

        <\/p>\n

        \u56fe 12\uff1a\u901a\u7528\u7684\u6570\u5b66\u8bc1\u660e\u751f\u6210\u5668 (LLM 1) \u548c\u9a8c\u8bc1\u5668 (LLM 2) \u8bbe\u7f6e\u3002<\/p>\n

        \u9a8c\u8bc1\u5668 LLM (LLM 2) \u63a5\u6536\u4e00\u4e2a\u8bc4\u5206\u6807\u51c6\u6765\u5bf9\u751f\u6210\u7684\u8bc1\u660e\u8fdb\u884c\u8bc4\u5206\uff0c\u5206\u6570\u89c4\u5219\u5982\u4e0b\uff1a<\/p>\n

          \n
        • \u300c1 \u5206\uff1a\u5b8c\u6574\u4e14\u4e25\u8c28\u7684\u8bc1\u660e\uff0c\u6240\u6709\u903b\u8f91\u6b65\u9aa4\u90fd\u6709\u6e05\u6670\u7684\u8bba\u8bc1\uff1b\u300d<\/li>\n
        • \u300c0.5 \u5206\uff1a\u8bc1\u660e\u6574\u4f53\u903b\u8f91\u5408\u7406\uff0c\u4f46\u6709\u5fae\u5c0f\u9519\u8bef\u6216\u9057\u6f0f\u7ec6\u8282\uff1b\u300d<\/li>\n
        • \u300c0 \u5206\uff1a\u8bc1\u660e\u5b58\u5728\u6839\u672c\u6027\u7f3a\u9677\uff0c\u5305\u542b\u81f4\u547d\u7684\u903b\u8f91\u9519\u8bef\u6216\u5173\u952e\u7f3a\u5931\u3002\u300d<\/li>\n<\/ul>\n

          \u5bf9\u4e8e\u8bc1\u660e\u9a8c\u8bc1\u5668\u6a21\u578b\uff0c\u4ed6\u4eec\u4ece DeepSeek V3.2-Exp-SFT \u5f00\u59cb\uff0c\u8fd9\u662f\u4ed6\u4eec\u57fa\u4e8e DeepSeek V3.2-Exp \u901a\u8fc7\u5728\u63a8\u7406\u6570\u636e\uff08\u6570\u5b66\u548c\u4ee3\u7801\uff09\u4e0a\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\u800c\u521b\u5efa\u7684\u6a21\u578b\u3002\u7136\u540e\uff0c\u4ed6\u4eec\u4f7f\u7528\u683c\u5f0f\u5956\u52b1\uff08\u68c0\u67e5\u89e3\u51b3\u65b9\u6848\u662f\u5426\u7b26\u5408\u9884\u671f\u683c\u5f0f\uff09\u548c\u57fa\u4e8e\u9884\u6d4b\u5206\u6570\u4e0e\u5b9e\u9645\u5206\u6570\uff08\u7531\u4eba\u7c7b\u6570\u5b66\u4e13\u5bb6\u6807\u6ce8\uff09\u63a5\u8fd1\u7a0b\u5ea6\u7684\u5206\u6570\u5956\u52b1\uff0c\u5bf9\u8be5\u6a21\u578b\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u3002<\/p>\n

          \u8bc1\u660e\u9a8c\u8bc1\u5668 (LLM 2) \u7684\u76ee\u6807\u662f\u68c0\u67e5\u751f\u6210\u7684\u8bc1\u660e (LLM 1)\uff0c\u4f46\u8c01\u6765\u68c0\u67e5\u8bc1\u660e\u9a8c\u8bc1\u5668\u5462\uff1f\u4e3a\u4e86\u4f7f\u8bc1\u660e\u9a8c\u8bc1\u5668\u66f4\u52a0\u7a33\u5065\u5e76\u9632\u6b62\u5176\u4ea7\u751f\u5e7b\u89c9\u95ee\u9898\uff0c\u4ed6\u4eec\u5f00\u53d1\u4e86\u7b2c\u4e09\u4e2a LLM\uff0c\u5373\u5143\u9a8c\u8bc1\u5668\uff08Meta-verifier\uff09\u3002<\/p>\n

          <\/p>\n

          \u56fe 13\uff1a\u5143\u9a8c\u8bc1\u5668 (LLM 3) \u68c0\u67e5\u9a8c\u8bc1\u5668 (LLM 2) \u662f\u5426\u6b63\u786e\u8bc4\u4f30\u4e86\u751f\u6210\u5668 (LLM 1)\u3002<\/p>\n

          \u5143\u9a8c\u8bc1\u5668 (LLM 3) \u4e5f\u662f\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u5f00\u53d1\u7684\uff0c\u7c7b\u4f3c\u4e8e LLM 2\u3002\u867d\u7136\u4f7f\u7528\u5143\u9a8c\u8bc1\u5668\u4e0d\u662f\u5fc5\u987b\u7684\uff0c\u4f46 DeepSeek \u56e2\u961f\u62a5\u544a\u79f0\uff1a<\/p>\n

          \u7531\u5143\u9a8c\u8bc1\u5668\u8bc4\u4f30\u7684\u9a8c\u8bc1\u5668\u8bc1\u660e\u5206\u6790\u7684\u5e73\u5747\u8d28\u91cf\u5f97\u5206\u4ece 0.85 \u63d0\u9ad8\u5230\u4e86 0.96\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u76f8\u540c\u7684\u8bc1\u660e\u8bc4\u5206\u9884\u6d4b\u51c6\u786e\u6027\u3002<\/p>\n

          \u8fd9\u5b9e\u9645\u4e0a\u662f\u4e00\u4e2a\u76f8\u5f53\u6709\u8da3\u7684\u8bbe\u7f6e\u3002\u5982\u679c\u4f60\u719f\u6089\u751f\u6210\u5bf9\u6297\u7f51\u7edc (GAN)\uff0c\u4f60\u53ef\u80fd\u4f1a\u5728\u8fd9\u91cc\u770b\u5230\u7c7b\u6bd4\u3002\u4f8b\u5982\uff0c\u8bc1\u660e\u9a8c\u8bc1\u5668\uff08\u5c06\u5176\u89c6\u4e3a GAN \u5224\u522b\u5668\uff09\u6539\u8fdb\u4e86\u8bc1\u660e\u751f\u6210\u5668\uff0c\u800c\u8bc1\u660e\u751f\u6210\u5668\u751f\u6210\u4e86\u66f4\u597d\u7684\u8bc1\u660e\uff0c\u8fdb\u4e00\u6b65\u63a8\u52a8\u4e86\u8bc1\u660e\u9a8c\u8bc1\u5668\u7684\u8fdb\u6b65\u3002<\/p>\n

          \u5143\u5206\u6570\uff08Meta score\uff09\u7528\u4e8e\u9a8c\u8bc1\u5668 (LLM 2) \u548c\u751f\u6210\u5668 (LLM 1) \u7684\u8bad\u7ec3\u671f\u95f4\u3002\u5b83\u5e76\u4e0d\u7528\u4e8e\u63a8\u7406\u65f6\u7684\u81ea\u6211\u4fee\u6b63\u5faa\u73af\uff0c\u6211\u4eec\u5c06\u5728\u4e0b\u4e00\u8282\u8ba8\u8bba\u8fd9\u4e00\u70b9\u3002<\/p>\n

          5.2 \u81ea\u6211\u4fee\u6b63<\/strong><\/p>\n

          \u5728\u4e0a\u4e00\u8282\u4e2d\uff0c\u6211\u4eec\u8c08\u5230\u4e86\u81ea\u6211\u9a8c\u8bc1\uff0c\u5373\u5206\u6790\u89e3\u51b3\u65b9\u6848\u7684\u8d28\u91cf\u3002\u5176\u76ee\u7684\u662f\u4e3a\u4e86\u5b9e\u73b0\u81ea\u6211\u4fee\u6b63\uff0c\u8fd9\u610f\u5473\u7740 LLM \u53ef\u4ee5\u6839\u636e\u53cd\u9988\u91c7\u53d6\u884c\u52a8\u5e76\u4fee\u6539\u5176\u7b54\u6848\u3002<\/p>\n

          \u4f20\u7edf\u4e0a\uff0c\u5728\u81ea\u6211\u4fee\u6b63\u8fd9\u79cd\u6210\u719f\u4e14\u6d41\u884c\u7684\u63a8\u7406\u6269\u5c55\u6280\u672f\u4e2d\uff0c\u6211\u4eec\u4f1a\u4f7f\u7528\u540c\u4e00\u4e2a LLM \u6765\u751f\u6210\u89e3\u51b3\u65b9\u6848\u5e76\u5bf9\u5176\u8fdb\u884c\u9a8c\u8bc1\uff0c\u7136\u540e\u518d\u8fdb\u884c\u4fee\u6b63\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5728\u524d\u9762\u7684\u56fe 12 \u548c 13 \u4e2d\uff0cLLM 1 \u548c LLM 2 \u5c06\u662f\u540c\u4e00\u4e2a LLM\u3002\u56e0\u6b64\uff0c\u4f20\u7edf\u7684\u81ea\u6211\u4fee\u6b63\u8fc7\u7a0b\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n

          <\/p>\n

          \u56fe 14\uff1a\u4e00\u4e2a\u7ecf\u5178\u7684\u81ea\u6211\u4fee\u6b63\u8fed\u4ee3\uff0c\u540c\u4e00\u4e2a LLM \u751f\u6210\u521d\u59cb\u54cd\u5e94\uff08Output 1\uff09\uff0c\u5bf9\u5176\u8fdb\u884c\u8bc4\u4f30\uff08Eval\uff09\uff0c\u5e76\u751f\u6210\u4fee\u6b63\u540e\u7684\u7b54\u6848\uff08Output 2\uff09\u3002<\/p>\n

          \u7136\u800c\uff0cDeepSeek \u56e2\u961f\u89c2\u5bdf\u5230\u5728\u5b9e\u8df5\u4e2d\u4f7f\u7528\u540c\u4e00\u4e2a LLM \u65e2\u8fdb\u884c\u751f\u6210\u53c8\u8fdb\u884c\u9a8c\u8bc1\u5b58\u5728\u4e00\u4e2a\u5173\u952e\u95ee\u9898\uff1a<\/p>\n

          \n

          \u5f53\u88ab\u63d0\u793a\u4e00\u6b21\u6027\u751f\u6210\u5e76\u5206\u6790\u5176\u81ea\u5df1\u7684\u8bc1\u660e\u65f6\uff0c\u751f\u6210\u5668\u5f80\u5f80\u4f1a\u58f0\u79f0\u6b63\u786e\uff0c\u5373\u4f7f\u5916\u90e8\u9a8c\u8bc1\u5668\u5f88\u5bb9\u6613\u53d1\u73b0\u7f3a\u9677\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u867d\u7136\u751f\u6210\u5668\u53ef\u4ee5\u6839\u636e\u5916\u90e8\u53cd\u9988\u4fee\u6b63\u8bc1\u660e\uff0c\u4f46\u5b83\u65e0\u6cd5\u50cf\u4e13\u7528\u9a8c\u8bc1\u5668\u90a3\u6837\u4e25\u8c28\u5730\u8bc4\u4f30\u81ea\u5df1\u7684\u5de5\u4f5c\u3002<\/p>\n<\/blockquote>\n

          \u4f5c\u4e3a\u903b\u8f91\u63a8\u8bba\uff0c\u4eba\u4eec\u4f1a\u5047\u8bbe\u4ed6\u4eec\u4f7f\u7528\u72ec\u7acb\u7684\u8bc1\u660e\u751f\u6210\u5668 (LLM 1) \u548c\u8bc1\u660e\u9a8c\u8bc1\u5668 (LLM 2)\u3002\u56e0\u6b64\uff0c\u8fd9\u91cc\u4f7f\u7528\u7684\u81ea\u6211\u4fee\u6b63\u5faa\u73af\u53d8\u5f97\u7c7b\u4f3c\u4e8e\u4e0b\u56fe\u6240\u793a\u3002\u8bf7\u6ce8\u610f\uff0c\u6211\u4eec\u7701\u7565\u4e86 LLM 3\uff0c\u5b83\u4ec5\u5728\u5f00\u53d1\u9a8c\u8bc1\u5668 (LLM 2) \u671f\u95f4\u4f7f\u7528\u3002<\/p>\n

          <\/p>\n

          \u56fe 15\uff1a\u4f7f\u7528\u72ec\u7acb\u9a8c\u8bc1\u5668 LLM (LLM 2) \u7684\u81ea\u6211\u4fee\u6b63\u3002<\/p>\n

          \u7136\u800c\uff0c\u5728\u5b9e\u8df5\u4e2d\uff0c\u4e0e\u56fe 15 \u4e0d\u540c\u7684\u662f\uff0cDeepSeek \u56e2\u961f\u4f7f\u7528\u7684\u751f\u6210\u5668\u548c\u9a8c\u8bc1\u5668 LLM \u4e0e\u56fe 14 \u4e2d\u7684\u7ecf\u5178\u81ea\u6211\u4fee\u6b63\u5faa\u73af\u662f\u540c\u4e00\u4e2a\uff1a<\/p>\n

          \n

          \u300c\u6240\u6709\u5b9e\u9a8c\u90fd\u4f7f\u7528\u4e86\u4e00\u4e2a\u5355\u4e00\u6a21\u578b\uff0c\u5373\u6211\u4eec\u7684\u6700\u7ec8\u8bc1\u660e\u751f\u6210\u5668\uff0c\u5b83\u540c\u65f6\u6267\u884c\u8bc1\u660e\u751f\u6210\u548c\u9a8c\u8bc1\u3002\u300d<\/p>\n<\/blockquote>\n

          \u6362\u53e5\u8bdd\u8bf4\uff0c\u72ec\u7acb\u7684\u9a8c\u8bc1\u5668\u5bf9\u4e8e\u8bad\u7ec3\u662f\u5fc5\u4e0d\u53ef\u5c11\u7684\u2014\u2014\u53ef\u7528\u4e8e\u6539\u8fdb\u751f\u6210\u5668\uff0c\u4f46\u5728\u751f\u6210\u5668\u8db3\u591f\u5f3a\u5927\u4e4b\u540e\uff0c\u5728\u63a8\u7406\u671f\u95f4\u5c31\u4e0d\u518d\u4f7f\u7528\uff08\u6216\u4e0d\u9700\u8981\uff09\u5b83\u4e86\u3002\u4e0e\u7b80\u5355\u7684\u5355\u6a21\u578b\u81ea\u6211\u4fee\u6b63\u7684\u5173\u952e\u533a\u522b\u5728\u4e8e\uff0c\u6700\u7ec8\u7684\u8bc1\u660e\u8005\u662f\u5728\u66f4\u5f3a\u5927\u7684\u9a8c\u8bc1\u5668\u548c\u5143\u9a8c\u8bc1\u5668\u7684\u6307\u5bfc\u4e0b\u8bad\u7ec3\u51fa\u6765\u7684\uff0c\u56e0\u6b64\u5b83\u5b66\u4f1a\u4e86\u5c06\u8fd9\u4e9b\u8bc4\u5206\u6807\u51c6\u5e94\u7528\u4e8e\u81ea\u5df1\u7684\u8f93\u51fa\u3002<\/p>\n

          \u6b64\u5916\uff0c\u5728\u63a8\u7406\u671f\u95f4\u4f7f\u7528\u8fd9\u79cd\u5408\u4e8c\u4e3a\u4e00\u7684 DeepSeekMath V2 \u9a8c\u8bc1\u5668\u5728\u8d44\u6e90\u548c\u6210\u672c\u65b9\u9762\u4e5f\u662f\u6709\u5229\u7684\uff0c\u56e0\u4e3a\u5b83\u6bd4\u8fd0\u884c\u7b2c\u4e8c\u4e2a LLM \u8fdb\u884c\u8bc1\u660e\u9a8c\u8bc1\u589e\u52a0\u4e86\u66f4\u5c11\u7684\u590d\u6742\u6027\u548c\u8ba1\u7b97\u9700\u6c42\u3002<\/p>\n

          \u56de\u5230\u56fe 14 \u548c 15 \u4e2d\u5c55\u793a\u7684\u4e00\u822c\u81ea\u6211\u4fee\u6b63\u6982\u5ff5\uff0c\u8fd9\u4e24\u5f20\u56fe\u90fd\u663e\u793a\u4e86 2 \u6b21\u8fed\u4ee3\u7684\u81ea\u6211\u4fee\u6b63\uff08\u521d\u59cb\u8fed\u4ee3\u548c\u4fee\u6b63\u540e\u7684\u7b54\u6848\uff09\u3002\u5f53\u7136\uff0c\u6211\u4eec\u53ef\u4ee5\u5411\u6b64\u8fc7\u7a0b\u6dfb\u52a0\u66f4\u591a\u8fed\u4ee3\u3002\u8fd9\u662f\u4e00\u4e2a\u7ecf\u5178\u7684\u63a8\u7406\u6269\u5c55\u6743\u8861\uff1a\u6211\u4eec\u6dfb\u52a0\u7684\u8fed\u4ee3\u8d8a\u591a\uff0c\u751f\u6210\u7b54\u6848\u7684\u6210\u672c\u5c31\u8d8a\u9ad8\uff0c\u4f46\u6574\u4f53\u51c6\u786e\u6027\u4e5f\u4f1a\u8d8a\u9ad8\u3002<\/p>\n

          \u5728\u8bba\u6587\u4e2d\uff0cDeepSeek \u56e2\u961f\u4f7f\u7528\u4e86\u591a\u8fbe 8 \u6b21\u8fed\u4ee3\uff0c\u770b\u8d77\u6765\u51c6\u786e\u6027\u5c1a\u672a\u9971\u548c\u3002<\/p>\n

          <\/p>\n

          \u56fe 16\uff1a\u989d\u5916\u7684\u81ea\u6211\u4fee\u6b63\u8fed\u4ee3\u63d0\u9ad8\u4e86\u51c6\u786e\u6027\u3002\u6765\u81ea DeepSeekMath V2 \u8bba\u6587\uff0c\u5e76\u52a0\u4e0a\u4e86\u6ce8\u91ca\u3002Best@32 \u591a\u6570\u6295\u7968\u65b9\u6cd5\u4e5f\u88ab\u79f0\u4e3a\u81ea\u6211\u4e00\u81f4\u6027\uff08Self-consistency\uff09\u3002<\/p>\n

          6. DeepSeek V3.2 (2025 \u5e74 12 \u6708 1 \u65e5)<\/h4>\n

          \u6211\u4eec\u5728\u4e0a\u4e00\u8282\u82b1\u4e86\u8fd9\u4e48\u591a\u65f6\u95f4\u8ba8\u8bba DeepSeekMath V2 \u7684\u539f\u56e0\u662f\uff1a<\/p>\n

            \n
          • \u5b83\u662f\u4e00\u4e2a\u975e\u5e38\u6709\u8da3\u7684\u6982\u5ff5\u9a8c\u8bc1\uff0c\u901a\u8fc7\u81ea\u6211\u9a8c\u8bc1\u548c\u81ea\u6211\u4fee\u6b63\u6280\u672f\u8fdb\u4e00\u6b65\u63a8\u52a8\u4e86\u300c\u5e26\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60\u300d\uff08RLVR\uff09\u7406\u5ff5\uff1b<\/li>\n
          • \u81ea\u6211\u9a8c\u8bc1\u548c\u81ea\u6211\u4fee\u6b63\u6280\u672f\u4e5f\u88ab\u7528\u4e8e DeepSeek V3.2 \u4e2d\u3002<\/li>\n<\/ul>\n

            \u4f46\u5728\u6211\u4eec\u8fdb\u5165\u8fd9\u90e8\u5206\u4e4b\u524d\uff0c\u5148\u6765\u770b\u770b DeepSeek V3.2 \u7684\u603b\u4f53\u6982\u51b5\u3002\u8fd9\u4e2a\u6a21\u578b\u4e4b\u6240\u4ee5\u91cd\u8981\uff0c\u662f\u56e0\u4e3a\u5b83\u4e0e\u5f53\u524d\u7684\u65d7\u8230\u6a21\u578b\u76f8\u6bd4\u8868\u73b0\u975e\u5e38\u51fa\u8272\u3002<\/p>\n

            <\/p>\n

            \u56fe 17\uff1aDeepSeek V3.2 \u4e0e\u4e13\u6709\u65d7\u8230\u6a21\u578b\u4e4b\u95f4\u7684\u57fa\u51c6\u6d4b\u8bd5\u5bf9\u6bd4\u3002\u6765\u81ea DeepSeek V3.2 \u62a5\u544a\uff0c\u5e76\u52a0\u4e0a\u4e86\u6ce8\u91ca\u3002<\/p>\n

            \u4e0e DeepSeek \u7684\u5176\u4ed6\u51e0\u6b3e\u6a21\u578b\u7c7b\u4f3c\uff0cV3.2 \u4e5f\u9644\u5e26\u4e86\u4e00\u4efd\u5f88\u68d2\u7684\u6280\u672f\u62a5\u544a\uff0c\u6211\u5c06\u5728\u63a5\u4e0b\u6765\u7684\u7ae0\u8282\u4e2d\u8fdb\u884c\u8ba8\u8bba\u3002<\/p>\n

            6.1 DeepSeek V3.2 \u67b6\u6784<\/strong><\/p>\n

            \u5f53\u7136\uff0c\u8be5\u6a21\u578b\u7684\u4e3b\u8981\u52a8\u673a\u662f\u63d0\u9ad8\u6574\u4f53\u6a21\u578b\u6027\u80fd\u3002\u4f8b\u5982\uff0c\u50cf DeepSeekMath V2 \u4e00\u6837\uff0c\u5b83\u5728\u6570\u5b66\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u83b7\u5f97\u4e86\u91d1\u724c\u7ea7\u7684\u8868\u73b0\u3002\u7136\u800c\uff0c\u8be5\u6a21\u578b\u5728\u8bad\u7ec3\u65f6\u4e5f\u8003\u8651\u5230\u4e86\u5de5\u5177\u7684\u4f7f\u7528\uff0c\u5e76\u4e14\u5728\u5176\u4ed6\u4efb\u52a1\uff08\u4f8b\u5982\u4ee3\u7801\u548c\u667a\u80fd\u4f53\u4efb\u52a1\uff09\u4e0a\u4e5f\u8868\u73b0\u826f\u597d\u3002<\/p>\n

            \u540c\u65f6\uff0cDeepSeek \u56e2\u961f\u5c06\u8ba1\u7b97\u6548\u7387\u89c6\u4e3a\u4e00\u4e2a\u5de8\u5927\u7684\u9a71\u52a8\u56e0\u7d20\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u4ed6\u4eec\u4f7f\u7528\u4e86 V2 \u548c V3 \u4e2d\u7684\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b (MLA) \u673a\u5236\uff0c\u4ee5\u53ca\u4ed6\u4eec\u5728 V3.2 \u4e2d\u6dfb\u52a0\u7684 DeepSeek \u7a00\u758f\u6ce8\u610f\u529b (DSA) \u673a\u5236\u3002\u4e8b\u5b9e\u4e0a\uff0c\u8bba\u6587\u4e2d\u63d0\u5230\u300cDeepSeek-V3.2 \u4f7f\u7528\u4e86\u4e0e DeepSeek-V3.2-Exp \u5b8c\u5168\u76f8\u540c\u7684\u67b6\u6784\u300d\uff0c\u8fd9\u6211\u4eec\u5728\u524d\u9762\u7684\u7ae0\u8282\u4e2d\u5df2\u7ecf\u8ba8\u8bba\u8fc7\u4e86\u3002<\/p>\n

            <\/p>\n

            \u56fe 18\uff1aDeepSeek V3.2 \u67b6\u6784\u3002<\/p>\n

            \u6b63\u5982\u6211\u4e4b\u524d\u63d0\u5230\u7684\uff0cDeepSeek V3.2-Exp \u7684\u53d1\u5e03\u5f88\u53ef\u80fd\u662f\u4e3a\u4e86\u8ba9\u751f\u6001\u7cfb\u7edf\u548c\u63a8\u7406\u57fa\u7840\u8bbe\u65bd\u51c6\u5907\u597d\u6258\u7ba1\u521a\u521a\u53d1\u5e03\u7684 V3.2 \u6a21\u578b\u3002<\/p>\n

            <\/p>\n

            \u56fe 19\uff1a\u7531 DeepSeek \u7a00\u758f\u6ce8\u610f\u529b (DSA) \u5b9e\u73b0\u7684\u63a8\u7406\u6210\u672c\u8282\u7701\u3002\u6765\u81ea DeepSeek V3.2 \u62a5\u544a\uff0c\u5e76\u52a0\u4e0a\u4e86\u5e26\u6ce8\u91ca\u3002<\/p>\n

            \u7531\u4e8e\u67b6\u6784\u4e0e DeepSeek V3.2-Exp \u76f8\u540c\uff0c\u6709\u8da3\u7684\u7ec6\u8282\u5728\u4e8e\u8bad\u7ec3\u65b9\u6cd5\uff0c\u6211\u4eec\u5c06\u5728\u63a5\u4e0b\u6765\u7684\u7ae0\u8282\u4e2d\u8ba8\u8bba\u3002<\/p>\n

            6.2 \u5f3a\u5316\u5b66\u4e60\u66f4\u65b0<\/strong><\/p>\n

            \u603b\u7684\u6765\u8bf4\uff0cDeepSeek \u56e2\u961f\u91c7\u7528\u4e86\u7c7b\u4f3c\u4e8e DeepSeek R1 \u7684\u300c\u5e26\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60\u300d\uff08RLVR\uff09\u7a0b\u5e8f\uff0c\u4f7f\u7528\u4e86\u7fa4\u4f53\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff08GRPO\uff09\u7b97\u6cd5\u3002\u4f46\u662f\uff0c\u6709\u4e00\u4e9b\u6709\u8da3\u7684\u66f4\u65b0\u503c\u5f97\u8ba8\u8bba\u3002<\/p>\n

            \u6700\u521d\uff0cDeepSeek R1 \u4f7f\u7528\u4e86\uff1a<\/p>\n

              \n
            • \u683c\u5f0f\u5956\u52b1\uff08\u786e\u4fdd\u7b54\u6848\u683c\u5f0f\u6b63\u786e\uff09\uff1b<\/li>\n
            • \u8bed\u8a00\u4e00\u81f4\u6027\u5956\u52b1\uff08\u786e\u4fdd\u6a21\u578b\u5728\u7f16\u5199\u56de\u590d\u65f6\u4e0d\u4f1a\u5728\u4e0d\u540c\u8bed\u8a00\u4e4b\u95f4\u5207\u6362\uff09\uff1b<\/li>\n
            • \u4e3b\u8981\u7684\u9a8c\u8bc1\u8005\u5956\u52b1\uff08\u6570\u5b66\u6216\u4ee3\u7801\u95ee\u9898\u4e2d\u7684\u7b54\u6848\u662f\u5426\u6b63\u786e\uff09\u3002<\/li>\n<\/ul>\n

              \u5bf9\u4e8e DeepSeek V3.2\uff0c\u4ed6\u4eec\u66f4\u6539\u4e86\u5956\u52b1\uff1a<\/p>\n

              \n

              \u5bf9\u4e8e\u63a8\u7406\u548c\u667a\u80fd\u4f53\u4efb\u52a1\uff0c\u6211\u4eec\u91c7\u7528\u57fa\u4e8e\u89c4\u5219\u7684\u7ed3\u679c\u5956\u52b1\u3001\u957f\u5ea6\u60e9\u7f5a\u548c\u8bed\u8a00\u4e00\u81f4\u6027\u5956\u52b1\u3002\u5bf9\u4e8e\u901a\u7528\u4efb\u52a1\uff0c\u6211\u4eec\u91c7\u7528\u751f\u6210\u5f0f\u5956\u52b1\u6a21\u578b\uff0c\u5176\u4e2d\u6bcf\u4e2a\u63d0\u793a\u90fd\u6709\u81ea\u5df1\u7684\u8bc4\u4f30\u6807\u51c6\u3002<\/p>\n<\/blockquote>\n

              \u4f8b\u5982\uff0c\u4ed6\u4eec\u79fb\u9664\u4e86\u683c\u5f0f\u5956\u52b1\uff0c\u4f46\u4e3a\u667a\u80fd\u4f53\u4efb\u52a1\u6dfb\u52a0\u4e86\u957f\u5ea6\u60e9\u7f5a\u3002\u7136\u540e\uff0c\u5bf9\u4e8e\u6ca1\u6709\u7b26\u53f7\u9a8c\u8bc1\u5668\uff08\u6570\u5b66\uff09\u6216\u4ee3\u7801\u89e3\u91ca\u5668\u6765\u9a8c\u8bc1\u7b54\u6848\u7684\u901a\u7528\u4efb\u52a1\uff0c\u4ed6\u4eec\u4f7f\u7528\u5956\u52b1\u6a21\u578b\uff08\u53e6\u4e00\u4e2a\u8bad\u7ec3\u7528\u4e8e\u8f93\u51fa\u5956\u52b1\u5206\u6570\u7684 LLM\uff09\u3002<\/p>\n

              \u6240\u4ee5\uff0c\u542c\u8d77\u6765\u8fd9\u4e2a\u6d41\u7a0b\u4e0d\u518d\u662f\u50cf DeepSeek R1 \u90a3\u6837\u7eaf\u7cb9\u7684\u57fa\u4e8e\u9a8c\u8bc1\u5668\u7684 RLVR\uff0c\u800c\u662f RLVR\uff08\u7528\u4e8e\u53ef\u9a8c\u8bc1\u9886\u57df\uff09\u548c\u66f4\u6807\u51c6\u7684\u300cLLM \u5373\u88c1\u5224\u300d\u5956\u52b1\u5efa\u6a21\uff08\u7528\u4e8e\u5176\u4ed6\u6240\u6709\u9886\u57df\uff09\u7684\u6df7\u5408\u4f53\u3002<\/p>\n

              \u5bf9\u4e8e\u6570\u5b66\u9886\u57df\uff0c\u4ed6\u4eec\u8868\u793a\u989d\u5916\u300c\u6574\u5408\u4e86\u6765\u81ea DeepSeekMath-V2 \u7684\u6570\u636e\u96c6\u548c\u5956\u52b1\u65b9\u6cd5\u300d\uff0c\u8fd9\u6211\u4eec\u5728\u672c\u6587\u524d\u9762\u5df2\u7ecf\u8ba8\u8bba\u8fc7\u4e86\u3002<\/p>\n

              6.3 GRPO \u66f4\u65b0<\/strong><\/p>\n

              \u5173\u4e8e RLVR \u6d41\u7a0b\u5185\u90e8\u7684\u5b66\u4e60\u7b97\u6cd5 GRPO \u672c\u8eab\uff0c\u81ea DeepSeek R1 \u8bba\u6587\u4e2d\u7684\u539f\u59cb\u7248\u672c\u4ee5\u6765\uff0c\u4ed6\u4eec\u4e5f\u505a\u4e86\u4e00\u4e9b\u66f4\u6539\u3002<\/p>\n

              \u5728\u8fc7\u53bb\u7684\u51e0\u4e2a\u6708\u91cc\uff0c\u6570\u5341\u7bc7\u8bba\u6587\u63d0\u51fa\u4e86\u5bf9 GRPO \u7684\u4fee\u6539\u5efa\u8bae\uff0c\u4ee5\u63d0\u9ad8\u5176\u7a33\u5b9a\u6027\u548c\u6548\u7387\u3002<\/p>\n

              \u5982\u679c\u4e0d\u6df1\u5165 GRPO \u7684\u6570\u5b66\u7ec6\u8282\uff0c\u7b80\u5355\u6765\u8bf4\uff0cDAPO \u4fee\u6539\u4e86 GRPO\uff0c\u91c7\u7528\u4e86\u975e\u5bf9\u79f0\u88c1\u526a\u3001\u52a8\u6001\u91c7\u6837\u3001Token \u7ea7\u635f\u5931\u548c\u663e\u5f0f\u7684\u57fa\u4e8e\u957f\u5ea6\u7684\u5956\u52b1\u6574\u5f62\u3002Dr. GRPO \u66f4\u6539\u4e86 GRPO \u76ee\u6807\u672c\u8eab\uff0c\u4ee5\u79fb\u9664\u957f\u5ea6\u548c\u6807\u51c6\u5dee\u5f52\u4e00\u5316\u3002<\/p>\n

              \u6700\u8fd1\u7684 Olmo 3 \u8bba\u6587\u4e5f\u91c7\u7528\u4e86\u7c7b\u4f3c\u7684\u53d8\u66f4\uff0c\u6211\u5f15\u7528\u5982\u4e0b\uff1a<\/p>\n

                \n
              • \u96f6\u68af\u5ea6\u4fe1\u53f7\u8fc7\u6ee4\uff1a \u6211\u4eec\u79fb\u9664\u5956\u52b1\u5b8c\u5168\u76f8\u540c\u7684\u5b9e\u4f8b\u7ec4\uff08\u5373\u4f18\u52bf\u6807\u51c6\u5dee\u4e3a\u96f6\u7684\u6279\u6b21\uff09\uff0c\u4ee5\u907f\u514d\u5728\u63d0\u4f9b\u96f6\u68af\u5ea6\u7684\u6837\u672c\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u7c7b\u4f3c\u4e8e DAPO (Yu et al., 2025)\u3002[DAPO]<\/li>\n
              • \u4e3b\u52a8\u91c7\u6837\uff1a \u5c3d\u7ba1\u8fdb\u884c\u4e86\u96f6\u68af\u5ea6\u8fc7\u6ee4\uff0c\u6211\u4eec\u4ecd\u4f7f\u7528\u4e00\u79cd\u65b0\u9896\u3001\u66f4\u9ad8\u6548\u7684\u52a8\u6001\u91c7\u6837\u7248\u672c\u7ef4\u6301\u4e00\u81f4\u7684\u6279\u6b21\u5927\u5c0f (Yu et al., 2025)\u3002\u8be6\u89c1 OlmoRL Infra\u3002[DAPO]<\/li>\n
              • Token \u7ea7\u635f\u5931\uff1a \u6211\u4eec\u4f7f\u7528 Token \u7ea7\u635f\u5931\uff0c\u901a\u8fc7\u6279\u6b21\u4e2d\u7684 Token \u603b\u6570\u8fdb\u884c\u5f52\u4e00\u5316 (Yu et al., 2025)\uff0c\u800c\u4e0d\u662f\u6309\u6837\u672c\u5f52\u4e00\u5316\uff0c\u4ee5\u907f\u514d\u957f\u5ea6\u504f\u5dee\u3002[DAPO]<\/li>\n
              • \u65e0 KL \u635f\u5931\uff1a \u4f5c\u4e3a\u4e00\u79cd\u5e38\u89c1\u505a\u6cd5\uff0c\u6211\u4eec\u79fb\u9664\u4e86 KL \u635f\u5931 (GLM-4.5 Team et al., 2025; Yu et al., 2025; Liu et al., 2025b)\uff0c\u56e0\u4e3a\u5b83\u5141\u8bb8\u66f4\u5c11\u9650\u5236\u7684\u7b56\u7565\u66f4\u65b0\uff0c\u5e76\u4e14\u79fb\u9664\u5b83\u4e0d\u4f1a\u5bfc\u81f4\u8fc7\u5ea6\u4f18\u5316\u6216\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u3002[DAPO \u548c Dr. GRPO]<\/li>\n
              • \u66f4\u9ad8\u88c1\u526a\u9608\u503c\uff1a \u6211\u4eec\u5c06\u635f\u5931\u4e2d\u7684\u4e0a\u9650\u88c1\u526a\u9879\u8bbe\u7f6e\u4e3a\u6bd4\u4e0b\u9650\u7565\u9ad8\u7684\u503c\uff0c\u4ee5\u5141\u8bb8\u5bf9 Token \u8fdb\u884c\u66f4\u5927\u7684\u66f4\u65b0\uff0c\u6b63\u5982 Yu et al. (2025) \u63d0\u8bae\u7684\u90a3\u6837\u3002[DAPO]<\/li>\n
              • \u622a\u65ad\u91cd\u8981\u6027\u91c7\u6837\uff1a \u4e3a\u4e86\u8c03\u6574\u63a8\u7406\u5f15\u64ce\u548c\u8bad\u7ec3\u5f15\u64ce\u4e4b\u95f4\u5bf9\u6570\u6982\u7387\u7684\u5dee\u5f02\uff0c\u6211\u4eec\u5c06\u635f\u5931\u4e58\u4ee5\u622a\u65ad\u7684\u91cd\u8981\u6027\u91c7\u6837\u6bd4\u7387\uff0c\u9075\u5faa Yao et al. (2025)\u3002<\/li>\n
              • \u65e0\u6807\u51c6\u5dee\u5f52\u4e00\u5316\uff1a \u5728\u8ba1\u7b97\u4f18\u52bf\u65f6\uff0c\u6211\u4eec\u4e0d\u5bf9\u7ec4\u7684\u6807\u51c6\u5dee\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u9075\u5faa Liu et al. (2025b)\u3002\u8fd9\u6d88\u9664\u4e86\u96be\u5ea6\u504f\u5dee\uff0c\u5373\u5956\u52b1\u6807\u51c6\u5dee\u4f4e\u7684\u95ee\u9898\uff08\u4f8b\u5982\u592a\u96be\u6216\u592a\u5bb9\u6613\uff09\u5176\u4f18\u52bf\u4f1a\u88ab\u5f52\u4e00\u5316\u9879\u663e\u8457\u653e\u5927\u3002[Dr. GRPO]<\/li>\n<\/ul>\n

                DeepSeek V3.2 \u4e2d\u7684 GRPO \u4fee\u6539\u7a0d\u5fae\u4e0d\u90a3\u4e48\u6fc0\u8fdb\uff0c\u6211\u7528\u7c7b\u4f3c\u4e8e Olmo 3 \u7684\u98ce\u683c\u603b\u7ed3\u5982\u4e0b\uff1a<\/p>\n

                  \n
                • \u7279\u5b9a\u9886\u57df\u7684 KL \u5f3a\u5ea6\uff08\u5305\u62ec\u6570\u5b66\u4e3a\u96f6\uff09\uff1a DeepSeek V3.2 \u6ca1\u6709\u50cf DAPO \u548c Dr. GRPO \u90a3\u6837\u5bf9\u6570\u5b66\u7c7b RL \u603b\u662f\u653e\u5f03 KL\uff0c\u800c\u662f\u5728\u76ee\u6807\u4e2d\u4fdd\u7559 KL \u9879\uff0c\u4f46\u6839\u636e\u6bcf\u4e2a\u9886\u57df\u8c03\u6574\u5176\u6743\u91cd\u3002\u7136\u800c\uff0c\u4ed6\u4eec\u4e5f\u6307\u51fa\uff0c\u975e\u5e38\u5f31\u751a\u81f3\u4e3a\u96f6\u7684 KL \u901a\u5e38\u5bf9\u6570\u5b66\u6548\u679c\u6700\u597d\u3002\uff08\u4f46\u4e0d\u662f\u5b8c\u5168\u79fb\u9664\u5b83\uff0c\u800c\u662f\u53d8\u6210\u4e86\u4e00\u4e2a\u8d85\u53c2\u6570\u3002\uff09<\/li>\n
                • \u65e0\u504f KL \u4f30\u8ba1\uff1a \u5982\u4e0a\u6240\u8ff0\uff0cDeepSeek V3.2 \u6ca1\u6709\u79fb\u9664 KL \u60e9\u7f5a\u3002\u9664\u4e86\u5c06\u5176\u89c6\u4e3a\u8c03\u8282\u65cb\u94ae\u5916\uff0c\u4ed6\u4eec\u8fd8\u63d0\u51fa\u4e86\u5bf9 GRPO \u4e2d KL \u60e9\u7f5a\u4f30\u8ba1\u65b9\u5f0f\u7684\u4fee\u6b63\uff0c\u5373\u7528\u7528\u4e8e\u4e3b\u635f\u5931\u7684\u76f8\u540c\u91cd\u8981\u6027\u6bd4\u7387\u91cd\u65b0\u52a0\u6743 KL \u9879\uff0c\u56e0\u6b64 KL \u68af\u5ea6\u5b9e\u9645\u4e0a\u4e0e\u6837\u672c\u6765\u81ea\u65e7\u7b56\u7565\u800c\u4e0d\u662f\u5f53\u524d\u7b56\u7565\u7684\u4e8b\u5b9e\u76f8\u5339\u914d\u3002<\/li>\n
                • \u5f02\u7b56\u7565\u5e8f\u5217\u63a9\u7801\uff08Off-policy sequence masking\uff09\uff1a \u5f53\u4ed6\u4eec\u5728\u8bb8\u591a\u68af\u5ea6\u6b65\u9aa4\u4e2d\u91cd\u7528 rollout \u6570\u636e\uff08rollout \u53ea\u662f\u6a21\u578b\u751f\u6210\u7684\u5b8c\u6574\u5e8f\u5217\u7684\u672f\u8bed\uff09\u65f6\uff0cDeepSeek V3.2 \u6d4b\u91cf\u5f53\u524d\u7b56\u7565\u5728\u6bcf\u4e2a\u5b8c\u6574\u7b54\u6848\u4e0a\u504f\u79bb rollout \u7b56\u7565\u7684\u7a0b\u5ea6\uff0c\u5e76\u7b80\u5355\u5730\u4e22\u5f03\u90a3\u4e9b\u65e2\u5177\u6709\u8d1f\u4f18\u52bf\u53c8\u300c\u8fc7\u4e8e\u504f\u79bb\u7b56\u7565\u300d\u7684\u5e8f\u5217\u3002\u56e0\u6b64\uff0c\u8fd9\u9632\u6b62\u4e86\u6a21\u578b\u4ece\u8fc7\u5ea6\u504f\u79bb\u7b56\u7565\u6216\u9648\u65e7\u7684\u6570\u636e\u4e2d\u5b66\u4e60\u3002<\/li>\n
                • \u4fdd\u7559 MoE \u6a21\u578b\u7684\u8def\u7531\uff1a \u5bf9\u4e8e\u6df7\u5408\u4e13\u5bb6\u9aa8\u5e72\u7f51\u7edc\uff0c\u4ed6\u4eec\u8bb0\u5f55\u4e86 rollout \u671f\u95f4\u6fc0\u6d3b\u4e86\u54ea\u4e9b\u4e13\u5bb6\uff0c\u5e76\u5728\u8bad\u7ec3\u671f\u95f4\u5f3a\u5236\u4f7f\u7528\u76f8\u540c\u7684\u8def\u7531\u6a21\u5f0f\uff0c\u4ee5\u4fbf\u68af\u5ea6\u66f4\u65b0\u9488\u5bf9\u90a3\u4e9b\u4ea7\u751f\u91c7\u6837\u7b54\u6848\u7684\u4e13\u5bb6\u3002<\/li>\n
                • \u4fdd\u7559 top-p \/top-k \u7684\u91c7\u6837\u63a9\u7801\uff1a \u5f53 rollout \u4f7f\u7528 top-p \u6216 top-k \u91c7\u6837\u65f6\uff0cDeepSeek V3.2 \u5b58\u50a8\u9009\u62e9\u63a9\u7801\u5e76\u5728\u8ba1\u7b97 GRPO \u635f\u5931\u548c KL \u65f6\u91cd\u65b0\u5e94\u7528\u5b83\uff0c\u4ee5\u4fbf\u8bad\u7ec3\u65f6\u7684\u52a8\u4f5c\u7a7a\u95f4\u4e0e\u91c7\u6837\u671f\u95f4\u5b9e\u9645\u53ef\u7528\u7684\u52a8\u4f5c\u7a7a\u95f4\u76f8\u5339\u914d\u3002<\/li>\n
                • \u4fdd\u7559\u539f\u59cb GRPO \u4f18\u52bf\u5f52\u4e00\u5316\uff1a Dr. GRPO \u8868\u660e GRPO \u7684\u957f\u5ea6\u548c\u6bcf\u7ec4\u6807\u51c6\u5dee\u5f52\u4e00\u5316\u9879\u4f1a\u4f7f\u4f18\u5316\u504f\u5411\u4e8e\u8fc7\u957f\u7684\u9519\u8bef\u7b54\u6848\uff0c\u5e76\u8fc7\u5ea6\u52a0\u6743\u975e\u5e38\u5bb9\u6613\u6216\u975e\u5e38\u96be\u7684\u95ee\u9898\u3002Dr. GRPO \u901a\u8fc7\u79fb\u9664\u8fd9\u4e24\u4e2a\u9879\u5e76\u56de\u5230\u65e0\u504f\u7684 PPO \u98ce\u683c\u76ee\u6807\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cDAPO \u8f6c\u5411 Token \u7ea7\u635f\u5931\uff0c\u8fd9\u4e5f\u6539\u53d8\u4e86\u957f\u7b54\u6848\u4e0e\u77ed\u7b54\u6848\u7684\u52a0\u6743\u65b9\u5f0f\u3002\u7136\u800c\uff0cDeepSeek V3.2 \u4fdd\u7559\u4e86\u539f\u59cb\u7684 GRPO \u5f52\u4e00\u5316\uff0c\u800c\u662f\u4e13\u6ce8\u4e8e\u5176\u4ed6\u4fee\u6b63\uff0c\u4f8b\u5982\u4e0a\u9762\u7684\u90a3\u4e9b\u3002<\/li>\n<\/ul>\n

                  \u6240\u4ee5\uff0c\u603b\u7684\u6765\u8bf4\uff0cDeepSeek V3.2 \u6bd4\u6700\u8fd1\u7684\u5176\u4ed6\u4e00\u4e9b\u6a21\u578b\u66f4\u63a5\u8fd1\u539f\u59cb\u7684 GRPO \u7b97\u6cd5\uff0c\u4f46\u589e\u52a0\u4e86\u4e00\u4e9b\u903b\u8f91\u4e0a\u7684\u5fae\u8c03\u3002<\/p>\n

                  6.4 DeepSeek V3.2-Speciale \u4e0e\u6269\u5c55\u601d\u7ef4<\/strong><\/p>\n

                  DeepSeek V3.2 \u8fd8\u6709\u4e00\u4e2a\u6781\u7aef\u7684\u3001\u6269\u5c55\u601d\u7ef4\uff08extended-thinking\uff09\u7684\u53d8\u4f53\uff0c\u79f0\u4e3a\u00a0DeepSeek V3.2-Speciale\uff0c\u5b83\u5728 RL \u9636\u6bb5\u4ec5\u5728\u63a8\u7406\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff08\u66f4\u7c7b\u4f3c\u4e8e DeepSeek R1\uff09\u3002\u9664\u4e86\u4ec5\u5728\u63a8\u7406\u6570\u636e\u4e0a\u8bad\u7ec3\u5916\uff0c\u4ed6\u4eec\u8fd8\u5728 RL \u671f\u95f4\u51cf\u5c11\u4e86\u957f\u5ea6\u60e9\u7f5a\uff0c\u5141\u8bb8\u6a21\u578b\u8f93\u51fa\u66f4\u957f\u7684\u54cd\u5e94\u3002<\/p>\n

                  \u751f\u6210\u66f4\u957f\u7684\u54cd\u5e94\u662f\u4e00\u79cd\u63a8\u7406\u6269\u5c55\u5f62\u5f0f\uff0c\u4e3a\u4e86\u83b7\u5f97\u66f4\u597d\u7684\u7ed3\u679c\uff0c\u54cd\u5e94\u56e0\u957f\u5ea6\u589e\u52a0\u800c\u53d8\u5f97\u66f4\u52a0\u6602\u8d35\u3002<\/p>\n

                  <\/p>\n

                  \u56fe 20\uff1a\u6269\u5c55\u601d\u7ef4\u7684 Speciale \u6a21\u578b\u5b9e\u73b0\u4e86\u66f4\u9ad8\u7684\u51c6\u786e\u6027\uff0c\u4f46\u4e5f\u751f\u6210\u4e86\u66f4\u591a\u7684 Token\u3002<\/p>\n

                  7. \u603b\u7ed3<\/h4>\n

                  \u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u6ca1\u6709\u6db5\u76d6 DeepSeek V3.2 \u8bad\u7ec3\u65b9\u6cd5\u7684\u6240\u6709\u7ec6\u8282\uff0c\u4f46\u6211\u5e0c\u671b\u4e0e\u4e4b\u524d\u7684 DeepSeek \u6a21\u578b\u7684\u6bd4\u8f83\u6709\u52a9\u4e8e\u9610\u660e\u4e3b\u8981\u89c2\u70b9\u548c\u521b\u65b0\u3002<\/p>\n

                  \u7b80\u800c\u8a00\u4e4b\uff0c\u6709\u8da3\u7684\u8981\u70b9\u662f\uff1a<\/p>\n

                    \n
                  • DeepSeek V3.2 \u4f7f\u7528\u4e86\u81ea DeepSeek V3 \u4ee5\u6765\u4e0e\u5176\u6240\u6709\u524d\u8eab\u76f8\u4f3c\u7684\u67b6\u6784\uff1b<\/li>\n
                  • \u4e3b\u8981\u7684\u67b6\u6784\u8c03\u6574\u662f\u4ed6\u4eec\u6dfb\u52a0\u4e86\u6765\u81ea DeepSeek V3.2-Exp \u7684\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\u4ee5\u63d0\u9ad8\u6548\u7387\uff1b<\/li>\n
                  • \u4e3a\u4e86\u63d0\u9ad8\u6570\u5b66\u6027\u80fd\uff0c\u4ed6\u4eec\u91c7\u7528\u4e86\u6765\u81ea DeepSeekMath V2 \u7684\u81ea\u6211\u9a8c\u8bc1\u65b9\u6cd5\uff1b<\/li>\n
                  • \u8bad\u7ec3\u6d41\u7a0b\u6709\u51e0\u9879\u6539\u8fdb\uff0c\u4f8b\u5982 GRPO \u7a33\u5b9a\u6027\u66f4\u65b0\uff08\u6ce8\u610f\u8bba\u6587\u8fd8\u6d89\u53ca\u56f4\u7ed5\u84b8\u998f\u3001\u957f\u4e0a\u4e0b\u6587\u8bad\u7ec3\u3001\u96c6\u6210\u7c7b\u4f3c\u4e8e gpt-oss \u7684\u5de5\u5177\u4f7f\u7528\u7b49\u5176\u4ed6\u51e0\u4e2a\u65b9\u9762\uff0c\u6211\u4eec\u5728\u672c\u6587\u4e2d\u6ca1\u6709\u6db5\u76d6\uff09\u3002<\/li>\n<\/ul>\n

                    \u65e0\u8bba DeepSeek \u6a21\u578b\u4e0e\u5176\u4ed6\u8f83\u5c0f\u7684\u5f00\u653e\u6743\u91cd\u6a21\u578b\u6216\u50cf GPT-5.1 \u6216 Gemini 3.0 Pro \u8fd9\u6837\u7684\u4e13\u6709\u6a21\u578b\u76f8\u6bd4\u7684\u5e02\u573a\u4efd\u989d\u5982\u4f55\uff0c\u6709\u4e00\u4ef6\u4e8b\u662f\u80af\u5b9a\u7684\uff1aDeepSeek \u7684\u53d1\u5e03\u603b\u662f\u5f88\u6709\u8da3\uff0c\u800c\u4e14\u4ece\u968f\u5f00\u653e\u6743\u91cd\u6a21\u578b\u68c0\u67e5\u70b9\u4e00\u8d77\u53d1\u5e03\u7684\u6280\u672f\u62a5\u544a\u4e2d\u603b\u6709\u5f88\u591a\u503c\u5f97\u5b66\u4e60\u7684\u4e1c\u897f\u3002<\/p>\n

                    \u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n

                     <\/p>\n<\/div>\n<\/div>\n

                    <\/div>\n

                    <\/g><\/svg><\/i> \"Loading\"<\/p>\n

                    <\/div>\n","protected":false},"excerpt":{"rendered":"

                    \u65e0\u8bba DeepSeek \u6a21\u578b\u4e0e\u5176\u4ed6\u8f83\u5c0f\u7684\u5f00\u653e\u6743\u91cd\u6a21\u578b\u6216\u50cf GPT-5.1 \u6216 Gemini 3.0 Pro \u8fd9 […]<\/p>\n

                    <\/div>\n

                    <\/g><\/svg><\/i> \"Loading\"<\/p>\n

                    <\/div>\n","protected":false},"author":56,"featured_media":28378,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23,20,80],"tags":[2013],"class_list":["post-28377","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23","category-20","category-80","tag-deepseek-v3v3-2"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/28377","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=28377"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/28377\/revisions"}],"predecessor-version":[{"id":28401,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/28377\/revisions\/28401"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/28378"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=28377"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=28377"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=28377"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}