{"id":31650,"date":"2026-09-08T16:37:11","date_gmt":"2026-09-08T08:37:11","guid":{"rendered":"http:\/\/139.9.1.231\/?p=31650"},"modified":"2026-09-08T16:37:16","modified_gmt":"2026-09-08T08:37:16","slug":"technical-report-on-automated-resource-management-for-large-model-training","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2026\/09\/08\/technical-report-on-automated-resource-management-for-large-model-training\/","title":{"rendered":"\u5927\u6a21\u578b\u81ea\u52a8\u5316\u8bad\u7ec3\u8d44\u6e90\u6280\u672f\u62a5\u544a"},"content":{"rendered":"\n<p class=\"has-light-pink-background-color has-background\"><strong><em>\u6458\u81ea\uff1aGithub <a href=\" https:\/\/github.com\/BinHPdev\/awesome-algorithm-auto-tools\"> https:\/\/github.com\/BinHPdev\/awesome-algorithm-auto-tools<\/a><\/em><\/strong><\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u6574\u7406\u65e5\u671f\uff1a2026-09-07\uff08\u5341\u6b21\u66f4\u65b0\u7248\uff09<br>\u6838\u5fc3\u65b9\u5411\uff1a<strong>\u7528 AI\/\u5927\u6a21\u578b \u81ea\u52a8\u5316\u9a71\u52a8\u6a21\u578b\u8bad\u7ec3\u3001\u5b9e\u9a8c\u3001\u8c03\u4f18\u7684\u5de5\u5177\u4e0e\u6846\u67b6<\/strong><br>\u53c2\u8003\u6807\u6746\uff1aKarpathy AutoResearch\u3001HuggingFace Skills\u3001AutoML \u7cfb\u5217<\/p><\/blockquote>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"\u76ee\u5f55\">\u76ee\u5f55<\/h2>\n\n\n\n<ol><li><strong><a href=\"#1-ai-\u81ea\u4e3b\u5b9e\u9a8c--\u7814\u7a76\u6846\u67b6\">AI \u81ea\u4e3b\u5b9e\u9a8c \/ \u7814\u7a76\u6846\u67b6<\/a><\/strong><\/li><li><strong><a href=\"#2-agent-\u9a71\u52a8\u7684\u8bad\u7ec3-skillshuggingface-\u751f\u6001\">Agent \u9a71\u52a8\u7684\u8bad\u7ec3 Skills\uff08HuggingFace \u751f\u6001\uff09<\/a><\/strong><\/li><li><strong><a href=\"#3-llm-\u5fae\u8c03\u6846\u67b6\u9ad8\u6548\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\">LLM \u5fae\u8c03\u6846\u67b6\uff08\u9ad8\u6548\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\uff09<\/a><\/strong><\/li><li><strong><a href=\"#4-\u5f3a\u5316\u5b66\u4e60\u5bf9\u9f50\u8bad\u7ec3\u6846\u67b6rlhf--grpo\">\u5f3a\u5316\u5b66\u4e60\u5bf9\u9f50\u8bad\u7ec3\u6846\u67b6\uff08RLHF \/ GRPO\uff09<\/a><\/strong><\/li><li><strong><a href=\"#5-\u81ea\u52a8\u5316\u8d85\u53c2\u6570\u4f18\u5316--automl-agent\">\u81ea\u52a8\u5316\u8d85\u53c2\u6570\u4f18\u5316 \/ AutoML Agent<\/a><\/strong><\/li><li><strong><a href=\"#6-\u81ea\u8fdb\u5316--self-play-\u8bad\u7ec3\u65b9\u6cd5\">\u81ea\u8fdb\u5316 \/ Self-Play \u8bad\u7ec3\u65b9\u6cd5<\/a><\/strong><\/li><li><strong><a href=\"#7-\u5408\u6210\u6570\u636e\u751f\u6210\u4e0e\u6570\u636e\u6cbb\u7406\">\u5408\u6210\u6570\u636e\u751f\u6210\u4e0e\u6570\u636e\u6cbb\u7406<\/a><\/strong><\/li><li><strong><a href=\"#8-\u77e5\u8bc6\u84b8\u998f\u6846\u67b6\">\u77e5\u8bc6\u84b8\u998f\u6846\u67b6<\/a><\/strong><\/li><li><strong><a href=\"#9-\u6a21\u578b\u5408\u5e76\u4e0e\u91cf\u5316\">\u6a21\u578b\u5408\u5e76\u4e0e\u91cf\u5316<\/a><\/strong><\/li><li><strong><a href=\"#10-\u8f7b\u91cf\u7ea7\u9884\u8bad\u7ec3\u4e0e\u5206\u5e03\u5f0f\u8bad\u7ec3\">\u8f7b\u91cf\u7ea7\u9884\u8bad\u7ec3\u4e0e\u5206\u5e03\u5f0f\u8bad\u7ec3<\/a><\/strong><\/li><li><strong><a href=\"#11-\u63a8\u7406\u5f15\u64cerl-\u8bad\u7ec3\u5faa\u73af\u6838\u5fc3\u7ec4\u4ef6\">\u63a8\u7406\u5f15\u64ce\uff08RL \u8bad\u7ec3\u5faa\u73af\u6838\u5fc3\u7ec4\u4ef6\uff09<\/a><\/strong><\/li><li><strong><a href=\"#12-\u591a\u6a21\u6001\u8bad\u7ec3\u6846\u67b6\">\u591a\u6a21\u6001\u8bad\u7ec3\u6846\u67b6<\/a><\/strong><\/li><li><strong><a href=\"#13-\u5b9e\u9a8c\u8ffd\u8e2a\u4e0e\u7f16\u6392\u5e73\u53f0\">\u5b9e\u9a8c\u8ffd\u8e2a\u4e0e\u7f16\u6392\u5e73\u53f0<\/a><\/strong><\/li><li><strong><a href=\"#14-benchmark-\u4e0e\u8bc4\u6d4b\u4f53\u7cfb\">Benchmark \u4e0e\u8bc4\u6d4b\u4f53\u7cfb<\/a><\/strong><\/li><li><strong><a href=\"#15-\u8f85\u52a9\u7f16\u7801-agent\u53ef\u7528\u4e8e\u8bad\u7ec3\u811a\u672c\u5f00\u53d1\">\u8f85\u52a9\u7f16\u7801 Agent\uff08\u53ef\u7528\u4e8e\u8bad\u7ec3\u811a\u672c\u5f00\u53d1\uff09<\/a><\/strong><\/li><li><strong><a href=\"#16-\u603b\u7ed3\u5bf9\u6bd4\u8868\u4e0e\u9009\u578b\u6307\u5357\">\u603b\u7ed3\u5bf9\u6bd4\u8868\u4e0e\u9009\u578b\u6307\u5357<\/a><\/strong><\/li><\/ol>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"1-ai-\u81ea\u4e3b\u5b9e\u9a8c--\u7814\u7a76\u6846\u67b6\">1. AI \u81ea\u4e3b\u5b9e\u9a8c \/ \u7814\u7a76\u6846\u67b6<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u8fd9\u4e00\u7c7b\u5de5\u5177\u7684\u6838\u5fc3\u662f\uff1a<strong>\u8ba9 AI Agent \u81ea\u4e3b\u8bbe\u8ba1\u5b9e\u9a8c\u3001\u4fee\u6539\u8bad\u7ec3\u4ee3\u7801\u3001\u8bc4\u4f30\u7ed3\u679c\u3001\u5faa\u73af\u8fed\u4ee3<\/strong>\u2014\u2014\u4eba\u7c7b\u7761\u89c9\uff0cAI \u505a\u5b9e\u9a8c\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"11-karpathy-autoresearch-\">1.1 Karpathy AutoResearch \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/karpathy\/autoresearch\">https:\/\/github.com\/karpathy\/autoresearch<\/a><\/li><li><strong>\u53d1\u5e03<\/strong>: 2026-03-06<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: AI Agent \u81ea\u4e3b\u8fdb\u884c ML \u5b9e\u9a8c\u5faa\u73af\uff1a\u8bfb\u4ee3\u7801 \u2192 \u63d0\u51fa\u5047\u8bbe \u2192 \u4fee\u6539\u4ee3\u7801 \u2192 \u8bad\u7ec3 5 \u5206\u949f \u2192 \u8bc4\u4f30 \u2192 \u4fdd\u7559\/\u4e22\u5f03 \u2192 \u5faa\u73af<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u4ec5 <strong>630 \u884c Python \u4ee3\u7801<\/strong>\uff0c\u6781\u7b80\u8bbe\u8ba1<\/li><li>\u56fa\u5b9a 5 \u5206\u949f\u65f6\u95f4\u9884\u7b97 \u2192 <strong>\u7ea6 12 \u6b21\u5b9e\u9a8c\/\u5c0f\u65f6\uff0c\u4e00\u665a\u7ea6 100 \u6b21\u5b9e\u9a8c<\/strong><\/li><li>\u57fa\u4e8e nanochat\uff08\u5355 GPU LLM \u8bad\u7ec3\u6846\u67b6\uff09<\/li><li><strong>\u5b9e\u9645\u6548\u679c<\/strong>: 2 \u5929\u7ea6 700 \u6b21\u81ea\u4e3b\u5b9e\u9a8c\uff0c\u53d1\u73b0\u7ea6 20 \u4e2a\u53ef\u53e0\u52a0\u6539\u8fdb\uff0c&#8221;Time to GPT-2&#8243; \u4ece 2.02h \u2192 1.80h\uff08<strong>11% \u6548\u7387\u63d0\u5347<\/strong>\uff09\uff0c\u4e14\u6539\u8fdb\u53ef\u8fc1\u79fb\u5230\u66f4\u5927\u6a21\u578b<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: ML \u6a21\u578b\u8bad\u7ec3\u8c03\u4f18\u3001\u8d85\u53c2\u6570\u641c\u7d22\u3001\u67b6\u6784\u63a2\u7d22\u3001\u8bad\u7ec3\u6280\u5de7\u53d1\u73b0<\/li><li><strong>\u8bb8\u53ef<\/strong>: MIT<\/li><li><strong>\u4e3a\u4ec0\u4e48\u91cd\u8981<\/strong>: \u8bc1\u660e\u4e86 &#8220;AI \u81ea\u4e3b\u505a ML \u7814\u7a76&#8221; \u7684\u8303\u5f0f\u53ef\u884c\uff0c\u662f\u672c\u62a5\u544a\u7684\u6838\u5fc3\u53c2\u8003\u8303\u5f0f<\/li><\/ul>\n\n\n\n<h3 id=\"12-sakana-ai-scientist-v2-\">1.2 Sakana AI Scientist v2 \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/SakanaAI\/AI-Scientist-v2\">https:\/\/github.com\/SakanaAI\/AI-Scientist-v2<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u5168\u81ea\u52a8\u5f00\u653e\u5f0f\u79d1\u5b66\u53d1\u73b0 \u2192 \u5047\u8bbe \u2192 \u5b9e\u9a8c\u8bbe\u8ba1 \u2192 \u6267\u884c \u2192 \u8bba\u6587\u64b0\u5199<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>Agentic Tree Search\uff1a\u6811\u72b6\u641c\u7d22\u5b9e\u9a8c\u7a7a\u95f4<\/li><li>VLM \u53cd\u9988\uff1a\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u7406\u89e3\u5b9e\u9a8c\u56fe\u8868<\/li><li><strong>\u5e76\u884c\u5b9e\u9a8c\u6267\u884c<\/strong><\/li><li>v2 \u6d88\u9664\u4e86\u5bf9\u4eba\u5de5\u4ee3\u7801\u6a21\u677f\u7684\u4f9d\u8d56\uff0c\u53ef\u8de8\u591a\u4e2a ML \u9886\u57df\u76f4\u63a5\u90e8\u7f72<\/li><\/ul><\/li><li><strong>\u76f8\u5173\u9879\u76ee<\/strong>:<ul><li><a href=\"https:\/\/github.com\/SakanaAI\">ShinkaEvolve<\/a> \u2014 \u7528 LLM \u4f5c\u4e3a\u53d8\u5f02\u7b97\u5b50\u7684\u7a0b\u5e8f\u8fdb\u5316\u6846\u67b6\uff0c\u7528\u4e8e\u79d1\u5b66\u53d1\u73b0<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u81ea\u52a8\u5316 ML \u7814\u7a76\u3001\u6a21\u578b\u67b6\u6784\u641c\u7d22\u3001\u8bad\u7ec3\u7b56\u7565\u63a2\u7d22<\/li><li><strong>\u8bb8\u53ef<\/strong>: \u5f00\u6e90<\/li><\/ul>\n\n\n\n<h3 id=\"13-automl-agent--\">1.3 AutoML-Agent \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/DeepAuto-AI\/automl-agent\">https:\/\/github.com\/DeepAuto-AI\/automl-agent<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: ICML 2025<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u591a Agent LLM \u6846\u67b6\u5b9e\u73b0\u5168\u6d41\u7a0b AutoML<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u5e76\u884c\u5316\u4e13\u7528 Agent\uff1a\u6570\u636e\u9884\u5904\u7406\u3001\u67b6\u6784\u8bbe\u8ba1\u3001\u8d85\u53c2\u4f18\u5316\u5404\u6709\u4e13\u95e8 Agent<\/li><li>\u68c0\u7d22\u589e\u5f3a\u89c4\u5212\uff08Retrieval-Augmented Planning\uff09<\/li><li>\u591a\u9636\u6bb5\u9a8c\u8bc1\u673a\u5236<\/li><li>\u5728 14 \u4e2a\u6570\u636e\u96c6\u4e0a\u6d4b\u8bd5\u9a8c\u8bc1<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5168\u81ea\u52a8\u5316 ML Pipeline<\/li><\/ul>\n\n\n\n<h3 id=\"14-auto-ml-agent-\">1.4 auto-ml-agent \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Nikhil-Doye\/auto-ml-agent\">https:\/\/github.com\/Nikhil-Doye\/auto-ml-agent<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: LLM \u7f16\u6392\u7684\u81ea\u4e3b ML Pipeline\uff0c\u8986\u76d6\u5b8c\u6574\u751f\u547d\u5468\u671f<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u4ece\u6570\u636e\u9884\u5904\u7406\u5230\u6a21\u578b\u90e8\u7f72\u5168\u81ea\u52a8\u5316\uff0c\u65e0\u9700\u4eba\u5de5\u5e72\u9884<\/li><li>\u591a Agent \u67b6\u6784 + \u6c99\u76d2\u6267\u884c<\/li><li>\u81ea\u7136\u8bed\u8a00\u62a5\u544a<\/li><li>Streamlit \u754c\u9762<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u7aef\u5230\u7aef\u81ea\u52a8\u5316\u6570\u636e\u79d1\u5b66\/ML \u5de5\u7a0b<\/li><\/ul>\n\n\n\n<h3 id=\"15-mlagentbench-\">1.5 MLAgentBench \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/snap-stanford\/MLAgentBench\">https:\/\/github.com\/snap-stanford\/MLAgentBench<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: Stanford SNAP Lab<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u8bc4\u6d4b AI Agent \u505a ML \u5b9e\u9a8c\u7684\u80fd\u529b<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>13 \u4e2a\u7aef\u5230\u7aef ML \u4efb\u52a1\uff08CIFAR-10 \u5230 BabyLM\uff09<\/li><li>Agent \u53ef\u6267\u884c\uff1a\u8bfb\u5199\u6587\u4ef6\u3001\u6267\u884c\u4ee3\u7801\u3001\u68c0\u67e5\u8f93\u51fa<\/li><li>Claude v3 Opus \u6700\u4f73\u6210\u529f\u7387 37.5%<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u8bc4\u4f30\/\u5f00\u53d1 ML \u7814\u7a76 Agent<\/li><\/ul>\n\n\n\n<h3 id=\"16-autoagent-\">1.6 AutoAgent \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/HKUDS\/AutoAgent\">https:\/\/github.com\/HKUDS\/AutoAgent<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u96f6\u4ee3\u7801\u521b\u5efa\u548c\u90e8\u7f72 LLM Agent<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u7528\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u9700\u6c42\u5373\u53ef\u751f\u6210 Agent<\/li><li>Self-play \u81ea\u6211\u5b9a\u5236\uff1a\u901a\u8fc7\u8fed\u4ee3\u81ea\u6211\u6539\u8fdb\u751f\u6210\u5de5\u5177\u3001Agent \u548c\u5de5\u4f5c\u6d41<\/li><li>\u652f\u6301\u81ea\u5b9a\u4e49\u5de5\u5177\u548c\u73af\u5883<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5feb\u901f\u642d\u5efa\u81ea\u52a8\u5316\u8bad\u7ec3 Agent<\/li><\/ul>\n\n\n\n<h3 id=\"17-ai-supervisor--\">1.7 AI-Supervisor \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2603.24402\">https:\/\/arxiv.org\/abs\/2603.24402<\/a><\/li><li><strong>\u53d1\u5e03<\/strong>: 2026-03-25<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u901a\u8fc7\u6301\u4e45\u5316\u7814\u7a76\u4e16\u754c\u6a21\u578b\uff08Knowledge Graph\uff09\u5b9e\u73b0\u81ea\u4e3b AI \u7814\u7a76\u76d1\u7763<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u591a Agent \u5171\u8bc6\u534f\u8bae\uff1a\u4e13\u7528 Agent \u5206\u5de5\u8fdb\u884c\u6587\u732e\u7efc\u8ff0\u3001\u5dee\u8ddd\u53d1\u73b0\u3001\u65b9\u6cd5\u5f00\u53d1\u3001\u8bc4\u4f30\u3001\u8bba\u6587\u64b0\u5199<\/li><li>\u6301\u7eed\u6f14\u5316\u7684 Research World Model\uff08\u77e5\u8bc6\u56fe\u8c31\uff09\uff0c\u8bb0\u5f55\u65b9\u6cd5\u3001\u57fa\u51c6\u3001\u5c40\u9650\u548c\u672a\u63a2\u7d22\u7a7a\u95f4<\/li><li>\u901a\u8fc7 GPU \u8ba1\u7b97\u548c API \u9a8c\u8bc1\u58f0\u660e\uff0c\u800c\u975e\u4ec5\u751f\u6210\u6587\u672c<\/li><li>\u6574\u5408 OpenReview \u7b49\u5e73\u53f0\u7684\u5ba1\u7a3f\u53cd\u9988<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5168\u81ea\u52a8\u5316 AI \u7814\u7a76\u76d1\u7763\u4e0e\u79d1\u5b66\u53d1\u73b0<\/li><\/ul>\n\n\n\n<h3 id=\"18-aris-auto-research-in-sleep--\">1.8 ARIS (Auto-Research-In-Sleep) \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/wanshuiyin\/Auto-claude-code-research-in-sleep\">https:\/\/github.com\/wanshuiyin\/Auto-claude-code-research-in-sleep<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u8f7b\u91cf\u7ea7 Markdown-only \u6280\u80fd\u7cfb\u7edf\uff0c\u8ba9 AI \u5728\u4f60\u7761\u89c9\u65f6\u505a ML \u7814\u7a76<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u96f6\u4f9d\u8d56\u3001\u96f6\u9501\u5b9a<\/strong>\uff1a\u6574\u4e2a\u7cfb\u7edf\u4ec5\u7531\u7eaf Markdown \u6587\u4ef6\u7ec4\u6210<\/li><li>\u8de8\u6a21\u578b\u5ba1\u67e5\u5faa\u73af\uff1aClaude Code \u505a\u7814\u7a76\uff0c\u5916\u90e8 LLM \u505a\u6279\u5224\u6027\u5ba1\u67e5<\/li><li>\u5b9e\u9645\u6548\u679c\uff1a\u4e00\u591c 4 \u8f6e\u5faa\u73af\u81ea\u4e3b\u8fd0\u884c 20+ GPU \u5b9e\u9a8c\uff0c\u91cd\u5199\u8bba\u6587\u53d9\u4e8b\uff0c\u5254\u9664\u7ad9\u4e0d\u4f4f\u811a\u7684\u58f0\u660e<\/li><li>\u517c\u5bb9\u4efb\u4f55 LLM Agent\uff08Codex CLI\u3001Cursor\u3001Windsurf \u7b49\uff09<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4e2a\u4eba\u7814\u7a76\u8005\u7684\u8fc7\u591c\u81ea\u52a8\u5316 ML \u7814\u7a76<\/li><\/ul>\n\n\n\n<h3 id=\"19-autokernel--\">1.9 AutoKernel \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/RightNow-AI\/autokernel\">https:\/\/github.com\/RightNow-AI\/autokernel<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u628a &#8220;Autoresearch&#8221; \u8303\u5f0f\u4e0b\u6c89\u5230 GPU kernel \u5c42\u2014\u2014\u7ed9\u5b83\u4efb\u610f PyTorch \u6a21\u578b\uff0c\u7761\u4e00\u89c9\uff0c\u9192\u6765\u5f97\u5230\u4f18\u5316\u540e\u7684 Triton kernel<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>1.2K+ stars\uff0c\u53d1\u5c55\u8fc5\u901f<\/li><li>\u5728 H100 \u4e0a\u5bf9 RMSNorm \u76f8\u6bd4 PyTorch eager \u5b9e\u73b0\u4e86 5.29x \u52a0\u901f<\/li><li>\u4e0e KernelBench \u96c6\u6210\uff0c\u81ea\u52a8\u8bc4\u6d4b\u751f\u6210\u7684 kernel<\/li><li>\u81ea\u4e3b Agent \u5faa\u73af\uff1aprofile \u2192 \u5408\u6210 \u2192 \u9a8c\u8bc1 \u2192 \u8fed\u4ee3<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\u4f18\u5316\uff1b\u4e3a\u81ea\u5b9a\u4e49\u7b97\u5b50\u5bfb\u627e\u9ad8\u6027\u80fd kernel \u5b9e\u73b0<\/li><\/ul>\n\n\n\n<h3 id=\"110-kernelagentmeta-pytorch-\">1.10 KernelAgent\uff08Meta PyTorch\uff09\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/meta-pytorch\/KernelAgent\">https:\/\/github.com\/meta-pytorch\/KernelAgent<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: Meta \u5b98\u65b9\u7684\u81ea\u4e3b GPU kernel \u751f\u6210\u4e0e\u4f18\u5316 Deep Agent \u6846\u67b6<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u96b6\u5c5e meta-pytorch \u5b98\u65b9\u7ec4\u7ec7<\/li><li>Auto-discover \u7b56\u7565\u81ea\u52a8\u63a2\u7d22 kernel \u4f18\u5316\u7a7a\u95f4<\/li><li>\u591a\u6a21\u578b\u4e2d\u7ee7\uff08multi-model relay\uff09\uff1a\u4e0d\u540c\u80fd\u529b\u7684 LLM \u5206\u5de5\u534f\u4f5c<\/li><li>2026-04 \u6301\u7eed\u6d3b\u8dc3\uff0cRMSNorm agent \u7b49\u91cd\u70b9\u7b97\u5b50\u91cd\u6784\u9646\u7eed\u843d\u5730<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5927\u5382\u7ea7\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd kernel \u81ea\u52a8\u5316\uff1b\u66ff\u4ee3\u624b\u5199 CUDA<\/li><\/ul>\n\n\n\n<h3 id=\"111-paperorchestra--\">1.11 PaperOrchestra \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Ar9av\/PaperOrchestra\">https:\/\/github.com\/Ar9av\/PaperOrchestra<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u57fa\u4e8e Google PaperOrchestra \u8bba\u6587\u7684\u5f00\u6e90\u5b9e\u73b0\u2014\u2014\u628a\u7c97\u60f3\u6cd5 + \u539f\u59cb\u5b9e\u9a8c\u65e5\u5fd7\uff0c\u81ea\u52a8\u8f6c\u6362\u6210\u53ef\u6295\u7a3f\u7ea7 LaTeX \u8bba\u6587<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u65e0\u9700 API Key\u3001\u65e0\u9700 LLM SDK\uff0c\u7eaf\u901a\u8fc7 Coding Agent \u7684 Skill \u8c03\u5ea6<\/li><li>\u652f\u6301 Claude Code\u3001Cursor\u3001Antigravity\u3001Cline\u3001Aider \u7b49\u4e3b\u6d41\u7f16\u7801 Agent<\/li><li>\u591a Agent \u534f\u4f5c\uff1a\u5199\u4f5c + \u6587\u732e\u7efc\u8ff0 + \u5f15\u7528\u9a8c\u8bc1 + \u914d\u56fe<\/li><li>\u5728 PaperWritingBench \u4e0a\u8d85\u8d8a\u5355 Agent \u4e0e\u6811\u641c\u7d22\u57fa\u7ebf<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: AI \u81ea\u4e3b\u7814\u7a76\u95ed\u73af\u7684\u6700\u540e\u4e00\u516c\u91cc\u2014\u2014\u4ece\u5b9e\u9a8c\u7ed3\u679c\u5230\u8bba\u6587\u4ea7\u51fa<\/li><\/ul>\n\n\n\n<h3 id=\"112-deep-researcher-agent--\">1.12 Deep Researcher Agent \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Xiangyue-Zhang\/auto-deep-researcher-24x7\">https:\/\/github.com\/Xiangyue-Zhang\/auto-deep-researcher-24&#215;7<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2604.05854\">https:\/\/arxiv.org\/abs\/2604.05854<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u4e1c\u4eac\u5927\u5b66\uff082026-04-07\uff09<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u8ba9 LLM Agent \u5168\u5929\u5019\uff0824\/7\uff09\u81ea\u4e3b\u8fd0\u884c\u6df1\u5ea6\u5b66\u4e60\u5b9e\u9a8c\uff0c\u8986\u76d6\u4ece\u5047\u8bbe\u63d0\u51fa\u5230\u7ed3\u679c\u5206\u6790\u7684\u5b8c\u6574\u5b9e\u9a8c\u751f\u547d\u5468\u671f<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u96f6\u6210\u672c\u76d1\u63a7<\/strong>\uff1a\u6a21\u578b\u8bad\u7ec3\u671f\u95f4\u4ec5\u4f9d\u8d56\u8fdb\u7a0b\u7ea7\u68c0\u67e5\u548c\u65e5\u5fd7\u8bfb\u53d6\uff0cLLM API \u8c03\u7528\u6210\u672c\u4e3a\u96f6<\/li><li><strong>\u53cc\u5c42\u6052\u5b9a\u5185\u5b58<\/strong>\uff1a\u4e0a\u9650\u7ea6 5K \u5b57\u7b26\uff0c\u65e0\u8bba\u8fd0\u884c\u591a\u957f\u65f6\u95f4\u5747\u4e0d\u589e\u957f\uff0c\u9632\u6b62\u4e0a\u4e0b\u6587\u7206\u70b8<\/li><li><strong>\u6700\u5c0f\u5de5\u5177\u96c6 Leader-Worker \u67b6\u6784<\/strong>\uff1a\u6bcf\u4e2a Worker Agent \u4ec5\u914d\u5907 3\u20135 \u4e2a\u5de5\u5177\uff0c\u6bcf\u6b21\u8c03\u7528 token \u5f00\u9500\u964d\u4f4e\u8fbe 73%<\/li><li><strong>\u5b9e\u9645\u6548\u679c<\/strong>\uff1a\u6301\u7eed\u90e8\u7f72 30+ \u5929\uff0c\u5728 4 \u4e2a\u5e76\u884c\u7814\u7a76\u9879\u76ee\u4e2d\u81ea\u4e3b\u5b8c\u6210 500+ \u5b9e\u9a8c\u5faa\u73af\uff1b\u5355\u4e2a\u9879\u76ee\u7ecf 200+ \u81ea\u52a8\u5b9e\u9a8c\u540e\u57fa\u7ebf\u63d0\u5347 52%<\/li><li>\u5e73\u5747 LLM \u6210\u672c\u4ec5 <strong>$0.08 \/ 24 \u5c0f\u65f6<\/strong><\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u8d44\u6e90\u53d7\u9650\u7684\u4e2a\u4eba\u7814\u7a76\u8005\uff1b\u9700\u8981\u957f\u671f\u65e0\u4eba\u503c\u5b88\u81ea\u52a8\u5316\u5b9e\u9a8c\u7684\u9879\u76ee<\/li><\/ul>\n\n\n\n<h3 id=\"113-ai-researcher\u6e2f\u5927-\">1.13 AI-Researcher\uff08\u6e2f\u5927\uff09\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/HKUDS\/AI-Researcher\">https:\/\/github.com\/HKUDS\/AI-Researcher<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2505.18705\">https:\/\/arxiv.org\/abs\/2505.18705<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u9999\u6e2f\u5927\u5b66\u6570\u636e\u667a\u80fd\u5b9e\u9a8c\u5ba4\uff08NeurIPS 2025 Spotlight\uff09<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u5168\u81ea\u52a8\u7814\u7a76\u7cfb\u7edf\u2014\u2014\u6587\u732e\u7efc\u8ff0 \u2192 \u5047\u8bbe\u751f\u6210 \u2192 \u7b97\u6cd5\u5b9e\u73b0 \u2192 \u8bba\u6587\u64b0\u5199\uff0c\u6700\u5c0f\u4eba\u5de5\u5e72\u9884<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>4.8K+ GitHub stars<\/strong>\uff0cNeurIPS 2025 Spotlight \u8bba\u6587<\/li><li><strong>Resource Analyst Agent<\/strong>\uff1a\u5c06\u7814\u7a76\u6982\u5ff5\u5206\u89e3\u4e3a\u539f\u5b50\u7ec4\u4ef6\uff0c\u5efa\u7acb\u6570\u5b66\u516c\u5f0f\u4e0e\u4ee3\u7801\u5b9e\u73b0\u7684\u53cc\u5411\u6620\u5c04\uff0c\u5927\u5e45\u964d\u4f4e\u5e7b\u89c9\u98ce\u9669<\/li><li>\u8fed\u4ee3\u7cbe\u70bc\u8303\u5f0f\uff1a\u4e13\u7528 Agent \u901a\u8fc7\u7ed3\u6784\u5316\u53cd\u9988\u5faa\u73af\u534f\u4f5c\uff08\u5199\u4f5c\u3001\u6587\u732e\u7efc\u8ff0\u3001\u5f15\u7528\u9a8c\u8bc1\u3001\u914d\u56fe\uff09<\/li><li>\u5c42\u6b21\u5316\u8bba\u6587\u5408\u6210\uff08Documentation Agent\uff09\uff0c\u89e3\u51b3\u957f\u6587\u672c\u8fde\u8d2f\u6027\u95ee\u9898<\/li><li><strong>Scientist-Bench<\/strong>\uff1a\u9996\u4e2a\u6807\u51c6\u5316\u81ea\u4e3b\u7814\u7a76\u8bc4\u6d4b\u57fa\u51c6\uff0c22 \u7bc7\u8bba\u6587\u3001\u591a\u4e2a AI \u9886\u57df<\/li><li>\u751f\u4ea7\u5c31\u7eea\uff1a\u5df2\u5728 Novix \u5e73\u53f0\uff08novix.science\uff09\u90e8\u7f72\u4e3a\u5546\u7528 AI Co-Scientist<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u81ea\u52a8\u5316 ML \u7814\u7a76\uff1b\u4ece\u5b9e\u9a8c\u5230\u8bba\u6587\u7684\u5b8c\u6574\u81ea\u52a8\u5316\u95ed\u73af<\/li><\/ul>\n\n\n\n<h3 id=\"114-deliautoresearch--\">1.14 Deli_AutoResearch \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>\u4f5c\u8005\/\u4e3b\u9875<\/strong>: Chen Deli\uff08\u9648\u5fb7\u529b\uff0cGitHub @victorchen96\uff09\u2014 <a href=\"https:\/\/victorchen96.github.io\/auto_research\/framework.html\">https:\/\/victorchen96.github.io\/auto_research\/framework.html<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u9762\u5411\u300c<strong>\u957f\u5468\u671f\uff08\u6570\u5929\u5230\u6570\u5468\uff09<\/strong>\u300d\u81ea\u4e3b\u4efb\u52a1\u7684\u534f\u8bae\u6846\u67b6\u2014\u2014\u53ea\u89c4\u5b9a\u7ea6\u5b9a\/\u89c4\u8303\u800c\u975e\u63d0\u4f9b\u53ef\u6267\u884c\u4ee3\u7801\uff0c\u4e13\u6cbb\u957f\u7a0b Agent \u7684\u4e09\u5927\u5931\u6548\u6a21\u5f0f\uff1a\u8ba4\u77e5\u6b7b\u5faa\u73af\u3001\u505c\u6ede\u3001\u8fd0\u884c\u65f6\u5d29\u6e83<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u7f16\u6392\u5668 + \u5168\u65b0\u4f1a\u8bdd Worker<\/strong>\uff1a\u6bcf\u4e2a\u4efb\u52a1\u5728\u9694\u79bb\u7684\u65b0\u4f1a\u8bdd\u4e2d\u6267\u884c\uff0c\u53ea\u4ece\u72b6\u6001\u6587\u4ef6\u8bfb\u53d6\u7cbe\u70bc\u4e0a\u4e0b\u6587\uff0c\u4ece\u4e0d\u9760\u5bf9\u8bdd\u5386\u53f2\u7eed\u8dd1 \u2192 \u675c\u7edd\u4e0a\u4e0b\u6587\u7d2f\u79ef<\/li><li><strong>\u9632\u5faa\u73af\u673a\u5236<\/strong>\uff1a\u5f3a\u5236\u65b9\u5411\u591a\u6837\u6027\uff0c<code>stale_count \u2265 2<\/code> \u65f6\u7ed3\u6784\u6027\u5207\u6362\u65b9\u5411<\/li><li><strong>\u4e09\u5c42\u5fc3\u8df3\u770b\u95e8\u72d7<\/strong>\uff1aShell \u5b88\u62a4\uff08L0\uff09+ \u6bcf\u5c0f\u65f6\u8c03\u5ea6\uff08L1\uff09+ \u4e1a\u52a1\u5faa\u73af\u56de\u8c03\uff08L2\uff09\uff0c\u68c0\u6d4b\u8d85\u8fc7 2 \u5c0f\u65f6\u7684\u505c\u6ede<\/li><li><strong>\u72b6\u6001\u6301\u4e45\u5316<\/strong>\uff1a<code>progress.json<\/code> \/ <code>findings.jsonl<\/code> \/ <code>directions_tried.json<\/code> \u7b49\u7248\u672c\u5316 JSON \u6587\u4ef6<\/li><li><strong>\u884c\u4e3a\u7ea6\u675f<\/strong>\uff1a\u8fd0\u884c\u671f\u95f4\u96f6\u7528\u6237\u4ea4\u4e92\u3001&#8221;ready \u5373\u6267\u884c&#8221;\uff08\u52a8\u4f5c\u524d\u4e0d\u4e8c\u6b21\u786e\u8ba4\uff09\u3001guardian\/worker \u89d2\u8272\u5206\u79bb<\/li><li><strong>\u5b9e\u6d4b<\/strong>\uff1a\u81ea\u4e3b\u4ea7\u51fa 4 \u7bc7\u7efc\u8ff0\u8bba\u6587\uff0859\u201375 \u9875 \/ \u5171 265 \u9875\u30011158 \u5f15\u7528\u3001\u81ea\u8bc4 8.5+\/10\uff09\uff0c\u542b\u4e00\u4e2a <strong>285B \u53c2\u6570 GRPO \u5b9e\u9a8c<\/strong>\uff1b\u6700\u957f\u5355\u6b21\u8fde\u7eed\u8fd0\u884c 72 \u5c0f\u65f6\uff0c\u7ea6 44h \u4ea7\u51fa\u5168\u90e8 4 \u7bc7<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u6570\u5929\/\u6570\u5468\u65e0\u4eba\u503c\u5b88\u7684\u957f\u7a0b\u81ea\u4e3b\u7814\u7a76\uff1b\u5f3a\u8c03\u300c\u6267\u884c\u4e0e\u8bc4\u4f30\u5206\u79bb + \u91cf\u5316\u505c\u6ede\u6307\u6807\u300d\u7684 Agent \u7f16\u6392\u8303\u5f0f<\/li><\/ul>\n\n\n\n<h3 id=\"115-autoresearchclaw--\">1.15 AutoResearchClaw \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/aiming-lab\/AutoResearchClaw\">https:\/\/github.com\/aiming-lab\/AutoResearchClaw<\/a><\/li><li><strong>\u4e00\u53e5\u8bdd<\/strong>: &#8220;Chat an Idea. Get a Paper. \ud83e\udd9e&#8221;\u2014\u2014\u5168\u81ea\u4e3b\u3001\u81ea\u8fdb\u5316\u7684\u300c\u60f3\u6cd5 \u2192 \u8bba\u6587\u300d\u7814\u7a76\u6d41\u6c34\u7ebf<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u8f93\u5165\u4e00\u4e2a\u7814\u7a76\u4e3b\u9898\uff0c\u7ecf <strong>23 \u9636\u6bb5\u6d41\u6c34\u7ebf<\/strong>\u81ea\u52a8\u4ea7\u51fa\u542b\u771f\u5b9e\u6587\u732e\u5f15\u7528\u3001\u6c99\u7bb1\u5b9e\u9a8c\u3001\u7edf\u8ba1\u5206\u6790\u3001\u540c\u884c\u8bc4\u5ba1\u3001\u4f1a\u8bae\u7ea7 LaTeX \u7684\u5b8c\u6574\u8bba\u6587<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>13.5K+ stars<\/strong>\uff0cMIT \u8bb8\u53ef\uff0c2026-05 \u4ecd\u6d3b\u8dc3\uff08v0.5.0 \u5f15\u5165\u591a\u9886\u57df\u5b9e\u9a8c Agent + ARC-Bench\uff09<\/li><li><strong>\u591a\u6e90\u6587\u732e\u53d1\u73b0<\/strong>\uff1aOpenAlex \/ Semantic Scholar \/ arXiv \u771f\u5b9e\u8bba\u6587\uff0c<strong>4 \u5c42\u5f15\u7528\u9a8c\u8bc1<\/strong>\u9632\u5e7b\u89c9<\/li><li><strong>\u786c\u4ef6\u611f\u77e5\u6267\u884c<\/strong>\uff1a\u81ea\u52a8\u8bc6\u522b NVIDIA GPU \/ Apple MPS \/ CPU\uff1bOpenCode Beast Mode \u751f\u6210\u590d\u6742\u5b9e\u9a8c<\/li><li><strong>6 \u79cd\u4eba\u673a\u534f\u4f5c\u6a21\u5f0f<\/strong>\uff1afull-auto \/ gate-only \/ checkpoint \/ step-by-step \/ co-pilot \/ custom\uff0c\u914d Idea Workshop\u3001Baseline Navigator\u3001Paper Co-Writer<\/li><li><strong>\u81ea\u5b66\u4e60<\/strong>\uff1a\u7ecf MetaClaw \u4ece\u6bcf\u6b21\u8fd0\u884c\u4e2d\u63d0\u70bc\u7ecf\u9a8c\uff1b\u53cd\u4f2a\u9020\u5b9e\u9a8c\u8bca\u65ad\u4e0e\u4fee\u590d\u5faa\u73af<\/li><li>\u5df2\u6f14\u793a\u5728 8 \u4e2a\u9886\u57df\uff08\u6570\u5b66\u3001\u7edf\u8ba1\u3001\u751f\u7269\u3001\u8ba1\u7b97\u3001NLP\u3001RL\u3001\u89c6\u89c9\u3001\u9c81\u68d2\u6027\uff09\u81ea\u52a8\u4ea7\u51fa 8 \u7bc7\u8bba\u6587<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u7aef\u5230\u7aef\u81ea\u4e3b\u79d1\u7814\uff1b\u4ece\u4e00\u4e2a idea \u76f4\u63a5\u4ea7\u51fa\u53ef\u6295\u7a3f\u8bba\u6587\uff0c\u53ef\u6309\u9700\u63d2\u5165\u4eba\u5de5\u5e72\u9884<\/li><\/ul>\n\n\n\n<h3 id=\"116-pi-autoresearch--\">1.16 pi-autoresearch \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/davebcn87\/pi-autoresearch\">https:\/\/github.com\/davebcn87\/pi-autoresearch<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u628a Karpathy \u7684 autoresearch \u5faa\u73af\u4ece\u300cML \u4e13\u7528\u300d\u6cdb\u5316\u4e3a<strong>\u9886\u57df\u65e0\u5173\u7684\u901a\u7528\u4f18\u5316\u539f\u8bed<\/strong>\u2014\u2014\u4efb\u4f55\u53ef\u91cf\u5316\u4e3a\u6570\u5b57\u7684\u6307\u6807\u90fd\u80fd\u81ea\u4e3b\u4f18\u5316<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>7K+ stars<\/strong>\uff0cMIT \u8bb8\u53ef\uff0c2026-06 \u6d3b\u8dc3<\/li><li><strong>\u6269\u5c55 = \u57fa\u7840\u8bbe\u65bd\uff0cSkill = \u9886\u57df\u77e5\u8bc6<\/strong>\uff1a\u4e00\u4e2a extension \u670d\u52a1\u65e0\u9650\u9886\u57df\uff0c\u53ea\u9700\u6362 skill<\/li><li>\u53ef\u4f18\u5316\u4efb\u610f\u53ef\u6d4b\u6307\u6807\uff1a\u6d4b\u8bd5\u6267\u884c\u65f6\u95f4\u3001JS bundle \u4f53\u79ef\u3001\u6784\u5efa\u901f\u5ea6\u3001Lighthouse \u5206\u6570\u3001LLM \u8bad\u7ec3 loss\u2026\u2026<\/li><li>\u300c\u8bd5\u60f3\u6cd5 \u2192 \u6d4b\u91cf \u2192 \u4fdd\u7559\u6709\u6548 \u2192 \u56de\u6eda\u56de\u9000 \u2192 \u6c38\u8fdc\u5faa\u73af\u300d\uff0c\u6bcf\u6b21\u5b9e\u9a8c\u8ffd\u52a0\u5199\u5165 JSONL + markdown \u4f1a\u8bdd\u6587\u6863<\/li><li><strong>\u53ef\u65ad\u70b9\u7eed\u8dd1<\/strong>\uff1a\u8de8\u91cd\u542f\u3001\u8de8 context reset \u65e0\u7f1d\u6062\u590d<\/li><li>\u914d\u5957\u751f\u6001\uff1a<code>pi-autoresearch-studio<\/code>\uff08\u4eea\u8868\u76d8 + \u8ba1\u5212\u7f16\u8f91\u5668 + PR \u5de5\u4f5c\u6d41\u7f16\u6392\uff09<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4e0d\u9650\u4e8e ML \u7684\u5de5\u7a0b\u4f18\u5316\u95ed\u73af\u2014\u2014\u524d\u7aef\u6027\u80fd\u3001\u6784\u5efa\u63d0\u901f\u3001\u6d4b\u8bd5\u52a0\u901f\u7b49\u4efb\u4f55\u300c\u8f93\u51fa\u4e00\u4e2a\u6570\u5b57\u300d\u7684\u4efb\u52a1<\/li><\/ul>\n\n\n\n<h3 id=\"117-openscience--\">1.17 OpenScience \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/synthetic-sciences\/openscience\">https:\/\/github.com\/synthetic-sciences\/openscience<\/a><\/li><li><strong>\u5b98\u7f51\/\u6587\u6863<\/strong>: <a href=\"https:\/\/www.openscience.sh\/\">https:\/\/www.openscience.sh\/<\/a><\/li><li><strong>\u53d1\u5e03<\/strong>: 2026-07-05\uff082026-07-22 \u6301\u7eed\u66f4\u65b0\uff09<\/li><li><strong>\u6765\u6e90<\/strong>: Synthetic Sciences<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u6a21\u578b\u65e0\u5173\u7684\u5f00\u6e90\u300cAI \u79d1\u7814\u5de5\u4f5c\u53f0\u300d\u2014\u2014\u628a\u6574\u4e2a\u79d1\u7814\u95ed\u73af\u4ea4\u7ed9 Agent \u81ea\u4e3b\u8fd0\u884c\uff0c\u8986\u76d6\u673a\u5668\u5b66\u4e60\u3001\u751f\u7269\u3001\u7269\u7406\u3001\u5316\u5b66\u7b49\u9886\u57df\uff0c\u800c\u4e0d\u53ea\u662f\u6587\u732e\u9605\u8bfb<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u8dd1\u5b8c\u6574\u7814\u7a76\u5faa\u73af<\/strong>\uff1a\u6587\u732e\u8c03\u7814 \u2192 \u63d0\u51fa\u5047\u8bbe \u2192 \u7f16\u5199\u4ee3\u7801 \u2192 \u5728\u771f\u5b9e\u7b97\u529b\u4e0a\u6267\u884c\u5b9e\u9a8c \u2192 \u5206\u6790\u7ed3\u679c \u2192 \u64b0\u5199\u62a5\u544a<\/li><li><strong>\u6a21\u578b\u5b8c\u5168\u53ef\u5207\u6362<\/strong>\uff1a\u4efb\u610f\u6a21\u578b\u7686\u53ef\u7528\uff08Claude\u3001GPT\u3001Gemini\u3001GLM\u3001Kimi\u3001DeepSeek\u3001\u672c\u5730\u5fae\u8c03\u6a21\u578b\uff09\uff0c\u53ef\u6309\u8bf7\u6c42\u9010\u6b21\u5207\u6362<\/li><li>\u5185\u7f6e <strong>250+ \u53ef\u7f16\u8f91 Skills<\/strong> + <strong>30+ \u79d1\u5b66\u6570\u636e\u5e93<\/strong>\uff08UniProt\u3001PDB\u3001ChEMBL\u3001arXiv \u7b49\uff09\u4f5c\u4e3a Agent \u5de5\u5177<\/li><li><strong>\u672c\u5730\u4f18\u5148\u3001\u65e0\u9065\u6d4b<\/strong>\uff1aAgent \u5faa\u73af\u3001Skills\u3001\u6570\u636e\u5e93\u5de5\u5177\u4e0e\u5de5\u4f5c\u533a\u5168\u90e8\u5f00\u6e90\u5728\u4ed3\u5e93\u4e2d\uff0cAPI Key \u7559\u5728\u672c\u673a\u3001\u8bf7\u6c42\u76f4\u8fde\u63d0\u4f9b\u5546<\/li><li>Apache-2.0 \u8bb8\u53ef\uff0cTypeScript \u5b9e\u73b0\uff0c\u53ef\u81ea\u6258\u7ba1<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u628a\u300c\u8bfb\u6587\u732e \u2192 \u5047\u8bbe \u2192 \u5b9e\u9a8c \u2192 \u8bba\u6587\u300d\u5168\u94fe\u8def\u4ea4\u7ed9 Agent \u7684\u8de8\u5b66\u79d1\u81ea\u4e3b\u79d1\u7814\uff1bML \u4e4b\u5916\u7684\u79d1\u5b66\u53d1\u73b0\u95ed\u73af<\/li><\/ul>\n\n\n\n<h3 id=\"118-arbor\u4eba\u5927-nlpir-\">1.18 Arbor\uff08\u4eba\u5927 NLPIR\uff09\u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/RUC-NLPIR\/Arbor\">https:\/\/github.com\/RUC-NLPIR\/Arbor<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2606.11926\uff08Toward\">https:\/\/arxiv.org\/abs\/2606.11926\uff08Toward<\/a> Generalist Autonomous Research via Hypothesis-Tree Refinement\uff09<\/li><li><strong>\u9879\u76ee\u4e3b\u9875<\/strong>: <a href=\"https:\/\/ruc-nlpir.github.io\/Arbor\/\">https:\/\/ruc-nlpir.github.io\/Arbor\/<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u4e2d\u56fd\u4eba\u6c11\u5927\u5b66 NLPIR \u5b9e\u9a8c\u5ba4\uff082026-06\uff09<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u901a\u7528\u81ea\u4e3b\u7814\u7a76 Agent\u2014\u2014\u7528 <strong>HTR\uff08Hypothesis Tree Refinement\uff0c\u5047\u8bbe\u6811\u7cbe\u70bc\uff09<\/strong> \u628a\u81ea\u4e3b\u7814\u7a76\u4ece\u300c\u4e00\u4e32\u5b64\u7acb\u7684\u5c40\u90e8\u5c1d\u8bd5\u300d\u53d8\u6210\u300c\u8de8\u65f6\u95f4\u7d2f\u79ef\u7684\u8fc7\u7a0b\u300d<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u957f\u751f\u547d\u5468\u671f Coordinator + \u77ed\u751f\u547d\u5468\u671f Executor<\/strong>\uff1aCoordinator \u4f5c\u4e3a\u300c\u7814\u7a76\u603b\u76d1\u300d\u7ef4\u62a4 Idea Tree\u3001\u9a71\u52a8 arbor \u5faa\u73af\u3001\u6d3e\u53d1\u5b9e\u9a8c\uff1bExecutor \u4f5c\u4e3a\u300c\u7814\u7a76\u5de5\u7a0b\u5e08\u300d\u5728\u9694\u79bb\u7684 git worktree \u4e2d\u5fe0\u5b9e\u5b9e\u73b0\u4ee3\u7801\u6539\u52a8\u5e76\u8dd1\u5b9e\u9a8c<\/li><li><strong>\u6301\u4e45\u5316\u5047\u8bbe\u6811<\/strong>\uff1a\u6bcf\u4e2a\u8282\u70b9\u7ed1\u5b9a\u300c\u5047\u8bbe + \u5b9e\u73b0\u8be5\u5047\u8bbe\u7684\u4ea7\u7269\u7248\u672c + \u5b9e\u9a8c\u8bc1\u636e + \u63d0\u70bc\u51fa\u7684\u6d1e\u89c1\u300d\uff0c\u7ed3\u679c\u3001\u5931\u8d25\u6a21\u5f0f\u4e0e\u6d1e\u89c1\u5411\u4e0a\u4f20\u64ad\uff0c\u540e\u7eed\u60f3\u6cd5\u8d77\u70b9\u66f4\u9ad8\uff0c\u800c\u4e0d\u662f\u968f\u4e0a\u4e0b\u6587\u6eda\u52a8\u4e22\u5931<\/li><li><strong>AO\uff08Autonomous Optimization\uff0c\u81ea\u4e3b\u4f18\u5316\uff09\u8bbe\u5b9a<\/strong>\uff1a\u4e0d\u7ed1\u5b9a\u7279\u5b9a benchmark\uff0c\u53ea\u8981\u300c\u6709\u53ef\u4f18\u5316\u7684\u4ea7\u7269 + \u660e\u786e\u76ee\u6807 + \u53ef\u6267\u884c\u53cd\u9988\u4fe1\u53f7\u300d\u5373\u53ef\u957f\u7a0b\u641c\u7d22\u2014\u2014\u7edf\u4e00\u8986\u76d6\u6a21\u578b\u8bad\u7ec3\u3001Harness \u5de5\u7a0b\u3001\u6570\u636e\u5408\u6210\u4e09\u7c7b\u7814\u7a76\u4efb\u52a1<\/li><li><strong>\u5b9e\u6d4b<\/strong>\uff1a6 \u4e2a\u771f\u5b9e\u7814\u7a76\u4efb\u52a1\u4e0a held-out \u7ed3\u679c\u5168\u90e8\u6700\u4f18\uff1b\u540c\u7b49\u8d44\u6e90\u9884\u7b97\u4e0b\u5e73\u5747\u76f8\u5bf9\u589e\u76ca\u8fbe Codex \u4e0e Claude Code \u7684 <strong>2.5 \u500d\u4ee5\u4e0a<\/strong>\uff1b\u540c\u4e00 Controller \u53ef\u76f4\u63a5\u8fc1\u79fb\u5230 MLE-Bench Lite<\/li><li>v0.1.0 \u53d1\u5e03\uff0c890+ stars<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u8de8\u5b9e\u9a8c\u7d2f\u79ef\u7ecf\u9a8c\u7684\u957f\u7a0b\u81ea\u4e3b\u7814\u7a76\uff1b\u6a21\u578b\u8bad\u7ec3\u8c03\u4f18\u3001Agent Harness \u5de5\u7a0b\u3001\u6570\u636e\u5408\u6210\u6d41\u7a0b\u7684\u81ea\u4e3b\u4f18\u5316<\/li><\/ul>\n\n\n\n<h3 id=\"119-sibyl-autoresearch--\">1.19 Sibyl-AutoResearch \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Sibyl-Research-Team\/AutoResearch-SibylSystem\">https:\/\/github.com\/Sibyl-Research-Team\/AutoResearch-SibylSystem<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2605.22343\uff08Autonomous\">https:\/\/arxiv.org\/abs\/2605.22343\uff08Autonomous<\/a> Research Needs Self-Evolving Trial-and-Error Harnesses, Not Paper Generators\uff09<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u8bba\u6587\u6807\u9898\u5373\u89c2\u70b9\u2014\u2014<strong>\u81ea\u4e3b\u7814\u7a76\u9700\u8981\u7684\u662f\u300c\u81ea\u8fdb\u5316\u7684\u8bd5\u9519 Harness\u300d\uff0c\u800c\u4e0d\u662f\u300c\u8bba\u6587\u751f\u6210\u5668\u300d<\/strong>\uff1bHarness \u8ba9 Agent \u8dd1\u6709\u754c\u8bd5\u9a8c\uff0c\u540c\u65f6\u4fdd\u7559\u6b63\u9762\u4e0e\u8d1f\u9762\u7ed3\u679c\uff0c\u5e76\u628a\u6559\u8bad\u56de\u6d41\u5230\u540e\u7eed\u89c4\u5212\u3001\u9a8c\u8bc1\u3001\u7ed3\u8bba\u8303\u56f4\u754c\u5b9a\u3001\u8c03\u5ea6\u3001\u6279\u5224\u4e0e\u5199\u4f5c\u4e2d<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u53cc\u5c42\u81ea\u8fdb\u5316<\/strong>\uff1a\u8fc7\u53bb\u7684\u8bd5\u9a8c\u6539\u53d8\u672a\u6765\u7684\u7814\u7a76\u884c\u4e3a\uff1b\u53cd\u590d\u51fa\u73b0\u7684<strong>\u6d41\u7a0b\u6027\u5931\u8d25<\/strong>\u5219\u6539\u53d8 Harness \u672c\u8eab<\/li><li><strong>20+ \u4e13\u7528 Agent + 19 \u9636\u6bb5\u72b6\u6001\u673a\u6d41\u6c34\u7ebf<\/strong>\uff1a\u6587\u732e\u8c03\u7814 \u2192 \u60f3\u6cd5\u751f\u6210 \u2192 \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u6267\u884c \u2192 \u7ed3\u679c\u5206\u6790 \u2192 \u8bba\u6587\u64b0\u5199 \u2192 \u540c\u884c\u8bc4\u5ba1\uff0c\u5168\u7a0b\u65e0\u9700\u4eba\u5de5\u4ecb\u5165<\/li><li><strong>Claude Code \u539f\u751f<\/strong>\uff08\u975e API \u5305\u88c5\uff09\uff1a\u76f4\u63a5\u6784\u5efa\u5728 fork skills\u3001agent teams\u3001MCP \u5de5\u5177\u4e4b\u4e0a\uff0c\u7ee7\u627f SSH \u8fdc\u7a0b\u6267\u884c\u3001\u591a\u6a21\u578b\u4ea4\u53c9\u8bc4\u5ba1\u3001\u4e91\u7aef\u540c\u6b65\u7b49\u751f\u6001\u80fd\u529b<\/li><li>GPU \u5b9e\u9a8c\u5e76\u884c\u8c03\u5ea6\uff08\u62d3\u6251\u6392\u5e8f\uff09\uff0c\u6587\u4ef6\u5316\u72b6\u6001\u5b58\u50a8<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u6709 GPU \u8d44\u6e90\u3001\u5e0c\u671b\u4ee5\u300c\u7ec4\u7ec7\u300d\u5f62\u6001\u8fd0\u884c\u81ea\u4e3b\u79d1\u7814\u7684\u56e2\u961f\uff1b\u5173\u6ce8 Harness \u81ea\u8eab\u53ef\u8fdb\u5316\u6027\u800c\u975e\u5355\u6b21\u8bba\u6587\u4ea7\u51fa\u7684\u7814\u7a76\u8005<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"2-agent-\u9a71\u52a8\u7684\u8bad\u7ec3-skillshuggingface-\u751f\u6001\">2. Agent \u9a71\u52a8\u7684\u8bad\u7ec3 Skills\uff08HuggingFace \u751f\u6001\uff09<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u8fd9\u4e00\u7c7b\u7684\u6838\u5fc3\u521b\u65b0\u662f\uff1a<strong>\u8ba9\u7f16\u7a0b Agent\uff08\u5982 Claude Code\uff09\u901a\u8fc7\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\u5b8c\u6210\u6a21\u578b\u8bad\u7ec3\u5168\u6d41\u7a0b<\/strong>\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"21-huggingface-skills-\">2.1 HuggingFace Skills \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/huggingface\/skills\">https:\/\/github.com\/huggingface\/skills<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u4e3a AI \u7f16\u7a0b Agent \u63d0\u4f9b\u6807\u51c6\u5316\u7684 ML \u4efb\u52a1 Skill \u5305\uff0c\u7528\u4e00\u53e5\u82f1\u6587\u5c31\u80fd\u5fae\u8c03\u6a21\u578b<\/li><li><strong>\u652f\u6301 Agent<\/strong>: Claude Code\u3001OpenAI Codex\u3001Google Gemini CLI\u3001Cursor<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>Skill \u540d\u79f0<\/th><th>\u529f\u80fd<\/th><th>\u5173\u952e\u6280\u672f<\/th><\/tr><\/thead><tbody><tr><td><strong>hugging-face-model-trainer<\/strong><\/td><td>\u5fae\u8c03\u8bed\u8a00\u6a21\u578b<\/td><td>TRL\uff08SFT\u3001DPO\u3001GRPO\uff09\uff0c0.5B-70B<\/td><\/tr><tr><td><strong>hugging-face-vision-trainer<\/strong><\/td><td>\u8bad\u7ec3\u89c6\u89c9\u6a21\u578b<\/td><td>RTDETRv2\u3001YOLOS\u3001DETR\u3001ViT<\/td><\/tr><tr><td><strong>hugging-face-jobs<\/strong><\/td><td>\u5728 HF \u57fa\u7840\u8bbe\u65bd\u4e0a\u8fd0\u884c\u8ba1\u7b97\u4efb\u52a1<\/td><td>\u4e91 GPU \u8c03\u5ea6\u3001\u6210\u672c\u4f30\u7b97<\/td><\/tr><tr><td><strong>hugging-face-trackio<\/strong><\/td><td>\u5b9e\u9a8c\u8ffd\u8e2a\u4e0e\u6307\u6807\u8bb0\u5f55<\/td><td>\u5b9e\u65f6 metrics logging<\/td><\/tr><tr><td><strong>hugging-face-evaluation<\/strong><\/td><td>\u6a21\u578b\u8bc4\u4f30\u5e76\u66f4\u65b0 model card<\/td><td>lighteval<\/td><\/tr><tr><td><strong>hugging-face-datasets<\/strong><\/td><td>\u521b\u5efa\/\u7ba1\u7406\u6570\u636e\u96c6<\/td><td>Hub API<\/td><\/tr><tr><td><strong>hugging-face-dataset-viewer<\/strong><\/td><td>\u6570\u636e\u96c6\u67e5\u8be2\u4e0e\u63a2\u7d22<\/td><td>REST API<\/td><\/tr><tr><td><strong>hf-cli<\/strong><\/td><td>Hub \u64cd\u4f5c\uff08\u4e0a\u4f20\/\u4e0b\u8f7d\/\u7ba1\u7406\u4ed3\u5e93\uff09<\/td><td>CLI<\/td><\/tr><tr><td><strong>gradio<\/strong><\/td><td>\u6784\u5efa\u4ea4\u4e92\u5f0f Demo<\/td><td>Web UI<\/td><\/tr><tr><td><strong>hugging-face-tool-builder<\/strong><\/td><td>\u6784\u5efa\u53ef\u590d\u7528 API \u81ea\u52a8\u5316\u811a\u672c<\/td><td>Python<\/td><\/tr><tr><td><strong>hugging-face-paper-publisher<\/strong><\/td><td>\u53d1\u5e03\u7814\u7a76\u8bba\u6587\u5230 Hub<\/td><td>\u2014<\/td><\/tr><tr><td><strong>transformers-js<\/strong><\/td><td>\u5728 JS\/TS \u4e2d\u8fd0\u884c ML \u6a21\u578b<\/td><td>ONNX Runtime<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<ul><li><strong>\u5178\u578b\u5de5\u4f5c\u6d41<\/strong>:<ol><li>\u7528\u81ea\u7136\u8bed\u8a00\u544a\u8bc9 Claude Code: &#8220;\u7528 GRPO \u5728\u8fd9\u4e2a\u6570\u636e\u96c6\u4e0a\u5fae\u8c03 Qwen-2.5-7B&#8221;<\/li><li>Agent \u81ea\u52a8\u751f\u6210\u8bad\u7ec3\u811a\u672c \u2192 \u63d0\u4ea4\u5230 HF Jobs \u2192 \u76d1\u63a7\u8bad\u7ec3\u6307\u6807 \u2192 \u8bc4\u4f30 checkpoint \u2192 \u63a8\u9001\u6a21\u578b\u5230 Hub<\/li><\/ol><\/li><li><strong>\u4e3a\u4ec0\u4e48\u91cd\u8981<\/strong>: \u8fd9\u662f &#8220;Vibe Training&#8221; \u7684\u8303\u5f0f\u2014\u2014\u7528\u81ea\u7136\u8bed\u8a00\u9a71\u52a8\u5b8c\u6574\u8bad\u7ec3\u6d41\u7a0b<\/li><\/ul>\n\n\n\n<h3 id=\"22-huggingface-autotrain-\">2.2 HuggingFace AutoTrain \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/huggingface\/autotrain-advanced\">https:\/\/github.com\/huggingface\/autotrain-advanced<\/a><\/li><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/huggingface.co\/autotrain\">https:\/\/huggingface.co\/autotrain<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2410.15735\">https:\/\/arxiv.org\/abs\/2410.15735<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: <strong>No-code<\/strong> \u6a21\u578b\u8bad\u7ec3\u5e73\u53f0\u2014\u2014\u4e0a\u4f20\u6570\u636e\uff0c\u9009\u62e9\u4efb\u52a1\uff0c\u81ea\u52a8\u5b8c\u6210\u4e00\u5207<\/li><li><strong>\u652f\u6301\u4efb\u52a1<\/strong>:<ul><li>LLM \u5fae\u8c03\uff08SFT\uff09<\/li><li>VLM \u5fae\u8c03<\/li><li>\u6587\u672c\u5206\u7c7b\/\u56de\u5f52<\/li><li>Token \u5206\u7c7b<\/li><li>Seq2Seq<\/li><li>Sentence Transformers \u5fae\u8c03<\/li><li>\u56fe\u50cf\u5206\u7c7b\/\u56de\u5f52<\/li><li>\u8868\u683c\u6570\u636e\u5206\u7c7b\/\u56de\u5f52<\/li><\/ul><\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u81ea\u52a8\u9009\u62e9\u6700\u4f18\u6a21\u578b<\/li><li>\u81ea\u52a8\u5904\u7406\u8bad\u7ec3\u3001\u8bc4\u4f30\u3001Hub \u53d1\u5e03<\/li><li>\u514d\u8d39\u4f7f\u7528\uff08\u4ec5\u4ed8\u8ba1\u7b97\u8d44\u6e90\u8d39\uff09<\/li><li>\u652f\u6301\u672c\u5730\u8fd0\u884c\u6216 HF Spaces<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5feb\u901f\u539f\u578b\u9a8c\u8bc1\u3001\u975e ML \u4e13\u5bb6\u7684\u6a21\u578b\u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"23-ml-intern--\">2.3 ml-intern \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/huggingface\/ml-intern\">https:\/\/github.com\/huggingface\/ml-intern<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: HuggingFace \u5b98\u65b9\u5f00\u6e90 &#8220;ML \u5de5\u7a0b\u5e08\u5b9e\u4e60\u751f&#8221;\u2014\u2014\u8bfb\u8bba\u6587\u3001\u5b9e\u73b0\u60f3\u6cd5\u3001\u5728 HF Jobs \u4e0a\u8dd1\u7aef\u5230\u7aef LLM \u540e\u8bad\u7ec3<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>900+ stars\uff0c2026-04-21 \u6b63\u5f0f\u516c\u5f00<\/li><li>\u57fa\u4e8e smolagents \u6784\u5efa\uff0c\u6df1\u5ea6\u6574\u5408 HF Hub \/ Jobs \u751f\u6001<\/li><li>\u63d0\u4f9b CLI \u4e0e Web UI \u4e24\u79cd\u4ea4\u4e92\u65b9\u5f0f<\/li><li>\u8986\u76d6\u5b8c\u6574\u540e\u8bad\u7ec3\u6d41\u7a0b\uff1aSFT\u3001DPO\u3001GRPO \u7b49<\/li><li>\u6bcf\u65e5\u6d3b\u8dc3\u63d0\u4ea4\uff0c\u8fed\u4ee3\u901f\u5ea6\u5feb<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u628a&#8221;\u60f3\u6cd5 \u2192 \u5b9e\u9a8c \u2192 \u843d\u5730\u6a21\u578b&#8221;\u5168\u94fe\u8def\u4ea4\u7ed9 Agent \u7684 HF \u7528\u6237<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"3-llm-\u5fae\u8c03\u6846\u67b6\u9ad8\u6548\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\">3. LLM \u5fae\u8c03\u6846\u67b6\uff08\u9ad8\u6548\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\uff09<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u8fd9\u4e9b\u662f\u5b9e\u9645\u6267\u884c\u8bad\u7ec3\u7684&#8221;\u5f15\u64ce&#8221;\u3002AutoResearch \/ HF Skills \u7b49\u4e0a\u5c42 Agent \u6700\u7ec8\u90fd\u9700\u8981\u8c03\u7528\u8fd9\u4e9b\u6846\u67b6\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"31-unsloth-\">3.1 Unsloth \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/unslothai\/unsloth\">https:\/\/github.com\/unslothai\/unsloth<\/a><\/li><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/unsloth.ai\">https:\/\/unsloth.ai<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: <strong>2x \u8bad\u7ec3\u901f\u5ea6\uff0c70% \u66f4\u5c11\u663e\u5b58<\/strong><\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u81ea\u5b9a\u4e49 CUDA kernel\uff0c\u5355 GPU \u573a\u666f\u65e0\u654c<\/li><li>\u652f\u6301 SFT + RL\uff08GRPO\u3001DPO\u3001PPO\uff09<\/li><li>FP8 \u5f3a\u5316\u5b66\u4e60\uff1a\u989d\u5916 1.4x \u52a0\u901f + 60% \u66f4\u5c11\u663e\u5b58<\/li><li>\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u8bad\u7ec3\uff1a\u4ec5 5GB VRAM \u5373\u53ef\u8bad\u7ec3\u63a8\u7406\u6a21\u578b<\/li><li>MoE \u6a21\u578b\u8bad\u7ec3\uff1a12x \u66f4\u5feb\uff0c35% \u66f4\u5c11 VRAM<\/li><li>\u652f\u6301 GGUF \u8f6c\u6362<\/li><li><strong>Unsloth MCP Server<\/strong>: <a href=\"https:\/\/github.com\/OtotaO\/unsloth-mcp-server\">https:\/\/github.com\/OtotaO\/unsloth-mcp-server<\/a> \u2014 \u53ef\u88ab AI Agent \u76f4\u63a5\u8c03\u7528<\/li><\/ul><\/li><li><strong>\u652f\u6301\u6a21\u578b<\/strong>: GPT-oss\u3001DeepSeek\u3001Qwen\u3001Llama\u3001Gemma \u7b49<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5355 GPU \u9ad8\u6548\u5fae\u8c03\u3001\u8d44\u6e90\u53d7\u9650\u73af\u5883<\/li><\/ul>\n\n\n\n<h3 id=\"32-axolotl-\">3.2 Axolotl \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u7075\u6d3b\u6027 + \u751f\u4ea7\u5c31\u7eea<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>YAML \u914d\u7f6e\u9a71\u52a8\uff0c\u6781\u6613\u4e0a\u624b<\/li><li>v0.8.x: QAT\uff08\u91cf\u5316\u611f\u77e5\u8bad\u7ec3\uff09\u3001\u5e8f\u5217\u5e76\u884c\uff08\u957f\u4e0a\u4e0b\u6587\uff09\u3001GRPO\uff08\u63a8\u7406\u8bad\u7ec3\uff09\u3001\u5b8c\u6574 RLHF Pipeline<\/li><li>\u793e\u533a\u6d3b\u8dc3\uff0c\u65b0\u6a21\u578b\/\u6280\u672f\u652f\u6301\u5feb<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u751f\u4ea7\u7ea7\u591a GPU \u5fae\u8c03<\/li><\/ul>\n\n\n\n<h3 id=\"33-llama-factory-llamafactory-\">3.3 LLaMA-Factory (LlamaFactory) \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/hiyouga\/LlamaFactory\">https:\/\/github.com\/hiyouga\/LlamaFactory<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: <strong>Web UI \u4e00\u7ad9\u5f0f\u5fae\u8c03<\/strong><\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>LlamaBoard Web UI\uff1a\u6d4f\u89c8\u5668\u4e2d\u914d\u7f6e\u548c\u542f\u52a8\u8bad\u7ec3<\/li><li>\u652f\u6301 100+ \u6a21\u578b<\/li><li>v0.9.4: OFT \u652f\u6301\u3001Megatron-LM \u96c6\u6210\u3001KTransformers \u540e\u7aef<\/li><li>\u652f\u6301 SFT\u3001RLHF\u3001DPO\u3001PPO \u7b49\u591a\u79cd\u8bad\u7ec3\u65b9\u6cd5<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5feb\u901f\u5b9e\u9a8c + \u4e0d\u60f3\u5199\u4ee3\u7801\u7684\u573a\u666f<\/li><\/ul>\n\n\n\n<h3 id=\"34-trl-transformer-reinforcement-learning-\">3.4 TRL (Transformer Reinforcement Learning) \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u6765\u6e90<\/strong>: HuggingFace \u5b98\u65b9<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: HuggingFace \u751f\u6001\u7684 RL \u8bad\u7ec3\u6807\u51c6\u5e93<\/li><li><strong>\u652f\u6301\u65b9\u6cd5<\/strong>: SFT\u3001DPO\u3001GRPO\u3001PPO\u3001KTO\u3001ORPO<\/li><li><strong>\u7279\u70b9<\/strong>: \u4e0e Transformers \/ PEFT \/ Accelerate \u6df1\u5ea6\u96c6\u6210<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5bf9\u9f50\u8bad\u7ec3\u3001\u63a8\u7406\u80fd\u529b\u8bad\u7ec3\uff08\u662f HF Skills model-trainer \u7684\u5e95\u5c42\u5f15\u64ce\uff09<\/li><\/ul>\n\n\n\n<h3 id=\"35-torchtune-\">3.5 torchtune \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u6765\u6e90<\/strong>: PyTorch \u5b98\u65b9\uff08Meta\uff09<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: PyTorch \u539f\u751f\uff0c\u65e0\u989d\u5916\u62bd\u8c61\u5c42<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>2025-02 \u652f\u6301\u591a\u8282\u70b9\u8bad\u7ec3<\/li><li>\u9ad8\u5ea6\u53ef\u6269\u5c55<\/li><li>\u4e0e PyTorch \u751f\u6001\u65e0\u7f1d\u96c6\u6210<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u6df1\u5ea6\u81ea\u5b9a\u4e49\u8bad\u7ec3\u6d41\u7a0b\u7684\u7814\u7a76\u8005<\/li><\/ul>\n\n\n\n<h3 id=\"36-nvidia-nemo-automodel-\">3.6 NVIDIA NeMo AutoModel \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/NVIDIA-NeMo\/Automodel\">https:\/\/github.com\/NVIDIA-NeMo\/Automodel<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: NVIDIA \u5b98\u65b9\uff0cDTensor-native SPMD\uff0cDay-0 HuggingFace \u652f\u6301<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u4ece\u5355\u8282\u70b9\u5230\u591a\u8282\u70b9\u591a GPU \u65e0\u7f1d\u6269\u5c55<\/li><li>\u652f\u6301 PEFT (LoRA) \u548c\u5168\u91cf SFT<\/li><li>\u6301\u7eed\u66f4\u65b0\u652f\u6301\u6700\u65b0\u6a21\u578b\uff08DeepSeek-V3.2\u3001Qwen3 \u7b49\uff09<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: NVIDIA GPU \u96c6\u7fa4\u4e0a\u7684\u5927\u89c4\u6a21\u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"37-lmflow--\">3.7 LMFlow \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/OptimalScale\/LMFlow\">https:\/\/github.com\/OptimalScale\/LMFlow<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u53ef\u6269\u5c55\u7684\u5927\u6a21\u578b\u5fae\u8c03\u4e0e\u63a8\u7406\u5de5\u5177\u7bb1<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>LISA \u5185\u5b58\u9ad8\u6548\u8bad\u7ec3<\/strong>\uff08\u6027\u80fd\u8d85\u8d8a LoRA\uff09<\/li><li>FlashAttention-1\/2 \u652f\u6301<\/li><li>NTK \u7f29\u653e\u652f\u6301\u957f\u4e0a\u4e0b\u6587<\/li><li>NAACL Best Demo Paper<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u5185\u5b58\u9ad8\u6548\u8bad\u7ec3\u7684\u7814\u7a76\u9879\u76ee<\/li><\/ul>\n\n\n\n<h3 id=\"38-h2o-llm-studio--\">3.8 H2O LLM Studio \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/h2oai\/h2o-llmstudio\">https:\/\/github.com\/h2oai\/h2o-llmstudio<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: <strong>\u65e0\u4ee3\u7801 GUI \u5fae\u8c03\u6846\u67b6<\/strong><\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u6d4f\u89c8\u5668\u7aef UI\uff0c\u65e0\u9700\u7f16\u5199\u4ee3\u7801<\/li><li>\u652f\u6301 LoRA \/ 4-bit \/ 8-bit \u91cf\u5316\u8bad\u7ec3<\/li><li>\u652f\u6301 DPO \/ IPO \/ KTO \u5bf9\u9f50\u65b9\u6cd5<\/li><li>\u56de\u5f52\u4e0e\u5206\u7c7b\u95ee\u9898\u7c7b\u578b<\/li><li>\u96c6\u6210 W&amp;B \/ Neptune \u5b9e\u9a8c\u8ffd\u8e2a<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u56e2\u961f\u5feb\u901f\u9a8c\u8bc1\u3001\u975e\u6280\u672f\u7528\u6237<\/li><\/ul>\n\n\n\n<h3 id=\"39-litgpt--\">3.9 LitGPT \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Lightning-AI\/litgpt\">https:\/\/github.com\/Lightning-AI\/litgpt<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: 20+ \u9ad8\u6027\u80fd LLM \u7684\u9884\u8bad\u7ec3\/\u5fae\u8c03\/\u90e8\u7f72\u4e00\u4f53\u5316<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>CLI \u9a71\u52a8\uff08<code>litgpt finetune\/pretrain\/evaluate<\/code>\uff09<\/li><li>4-bit &amp; 8-bit \u91cf\u5316\u652f\u6301<\/li><li>\u66fe\u9a71\u52a8 TinyLlama \u9879\u76ee<\/li><li>NeurIPS 2023 LLM Efficiency Challenge \u8d77\u59cb\u5957\u4ef6<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5feb\u901f\u539f\u578b + CLI \u504f\u597d\u7528\u6237<\/li><\/ul>\n\n\n\n<h3 id=\"310-instructlab--\">3.10 InstructLab \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/instructlab\">https:\/\/github.com\/instructlab<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: IBM \/ Red Hat<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u901a\u8fc7\u5408\u6210\u6570\u636e\u534f\u4f5c\u5b9a\u5236 LLM<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>LAB \u5bf9\u9f50\u65b9\u6cd5<\/li><li>\u57fa\u4e8e\u5206\u7c7b\u5b66\uff08Taxonomy\uff09\u7684\u6280\u80fd\u8d21\u732e\u673a\u5236<\/li><li>\u4ece\u5c11\u91cf\u4eba\u5de5\u79cd\u5b50\u6570\u636e\u751f\u6210\u5927\u89c4\u6a21\u8bad\u7ec3\u6570\u636e<\/li><li>\u9762\u5411 Granite \u6a21\u578b\u65cf<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u793e\u533a\u534f\u4f5c\u5f0f\u6a21\u578b\u6539\u8fdb<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"4-\u5f3a\u5316\u5b66\u4e60\u5bf9\u9f50\u8bad\u7ec3\u6846\u67b6rlhf--grpo\">4. \u5f3a\u5316\u5b66\u4e60\u5bf9\u9f50\u8bad\u7ec3\u6846\u67b6\uff08RLHF \/ GRPO\uff09<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>2025-2026 \u8d8b\u52bf\uff1a<strong>GRPO\uff08Group Relative Policy Optimization\uff09\u6b63\u5728\u53d6\u4ee3 PPO \u6210\u4e3a\u9ed8\u8ba4\u5bf9\u9f50\u65b9\u6cd5<\/strong>\u2014\u2014\u65e0\u9700 Critic \u6a21\u578b\uff0c\u66f4\u7b80\u5355\u7a33\u5b9a\u3002<br>2026 \u65b0\u8d8b\u52bf\uff1a<strong>\u5404\u5927\u5382\u548c\u5b9e\u9a8c\u5ba4\u7eb7\u7eb7\u5f00\u6e90\u81ea\u5df1\u7684 RL \u8bad\u7ec3\u6846\u67b6<\/strong>\uff0c\u5f62\u6210\u767e\u82b1\u9f50\u653e\u683c\u5c40\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"41-openrlhf-\">4.1 OpenRLHF \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/OpenRLHF\/OpenRLHF\">https:\/\/github.com\/OpenRLHF\/OpenRLHF<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u9ad8\u6027\u80fd RLHF \u6846\u67b6\uff0cRay + vLLM \u5206\u5e03\u5f0f\u67b6\u6784<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u652f\u6301 70B+ \u5168\u91cf\u5fae\u8c03<\/li><li>PPO\u3001DAPO\u3001REINFORCE++\u3001TIS<\/li><li>\u5f02\u6b65 RL \u8bad\u7ec3 (<code>--async_train<\/code>)<\/li><li>\u5f02\u6b65 Agent RLHF (<code>--agent_func_path<\/code>)<\/li><li><strong>MARTI \u5206\u652f<\/strong>: \u591a Agent RL \u8bad\u7ec3\u7cfb\u7edf<\/li><li><strong>OpenRLHF-M<\/strong>: \u591a\u6a21\u6001\u6a21\u578b RLHF<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5927\u89c4\u6a21 RLHF\/\u5bf9\u9f50\u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"42-verlvolcano-engine-rl-\">4.2 verl\uff08Volcano Engine RL\uff09\u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/volcengine\/verl\">https:\/\/github.com\/volcengine\/verl<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u5b57\u8282\u8df3\u52a8\u706b\u5c71\u5f15\u64ce<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u9ad8\u6027\u80fd\u3001\u529f\u80fd\u4e30\u5bcc\u7684 LLM RL \u8bad\u7ec3\u6808<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>HybridFlow \u67b6\u6784\u8bba\u6587<\/li><li>\u51e0\u884c\u4ee3\u7801\u5373\u53ef\u5b9e\u73b0 GRPO\/PPO<\/li><li>3D-HybridEngine\uff1a\u96c6\u6210 FSDP\u3001Megatron-LM\u3001vLLM\u3001SGLang<\/li><li><strong>\u4f7f\u7528\u65b9<\/strong>: ByteDance\u3001Alibaba Qwen\u3001Anyscale\u3001LMSys\u3001UC Berkeley<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u751f\u4ea7\u7ea7 RL \u8bad\u7ec3\uff0c\u5df2\u88ab\u5de5\u4e1a\u754c\u5e7f\u6cdb\u9a8c\u8bc1<\/li><\/ul>\n\n\n\n<h3 id=\"43-dapo--\">4.3 DAPO \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/BytedTsinghua-SIA\/DAPO\">https:\/\/github.com\/BytedTsinghua-SIA\/DAPO<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u5b57\u8282\u8df3\u52a8 Seed + \u6e05\u534e AIR<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u5f00\u6e90 RL \u7cfb\u7edf\uff0cAIME 2024 \u4e0a\u8fbe\u5230 50 \u5206<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>4 \u9879\u5173\u952e\u7a33\u5b9a\u6027\u6280\u672f\uff1aClip-Higher\u3001Dynamic Sampling\u3001Token-Level PG Loss\u3001Overlong Reward Shaping<\/li><li>\u57fa\u4e8e verl \u6784\u5efa<\/li><li>Qwen2.5-32B \u4e0a\u9a8c\u8bc1\u6548\u679c<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u6570\u5b66\u63a8\u7406\u80fd\u529b RL \u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"44-areal--\">4.4 AReaL \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/inclusionAI\/AReaL\">https:\/\/github.com\/inclusionAI\/AReaL<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u8682\u8681\u96c6\u56e2 + \u6e05\u534e\u5927\u5b66<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: <strong>\u5168\u5f02\u6b65 RL \u8bad\u7ec3\uff0c\u901f\u5ea6\u6700\u5feb<\/strong><\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>v0.3: \u76f8\u6bd4\u540c\u6b65\u8bad\u7ec3 <strong>2.77x \u52a0\u901f<\/strong><\/li><li>AReaL-lite \u7248\u672c\u51cf\u5c11 80% \u4ee3\u7801\u91cf<\/li><li>\u5347\u817e NPU \u652f\u6301<\/li><li>GSPO \u7b97\u6cd5<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u8ffd\u6c42\u6781\u81f4\u8bad\u7ec3\u901f\u5ea6\u7684\u573a\u666f<\/li><\/ul>\n\n\n\n<h3 id=\"45-slime--\">4.5 slime \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/THUDM\/slime\">https:\/\/github.com\/THUDM\/slime<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u6e05\u534e\u5927\u5b66 \/ GLM \u56e2\u961f<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u9a71\u52a8 GLM-4.5\/4.6\/4.7\/5 \u6a21\u578b\u8bad\u7ec3\u7684 RL \u6846\u67b6<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>Megatron + SGLang \u67b6\u6784<\/li><li>\u652f\u6301 Qwen3\u3001DeepSeek V3<\/li><li>RLVE\uff1a400 \u4e2a\u53ef\u9a8c\u8bc1\u73af\u5883<\/li><li>TritonForge GPU kernel \u8bad\u7ec3<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5927\u89c4\u6a21\u6a21\u578b\u7684 RL \u540e\u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"46-nemo-rl--\">4.6 NeMo RL \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/NVIDIA-NeMo\/RL\">https:\/\/github.com\/NVIDIA-NeMo\/RL<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: NVIDIA \u5b98\u65b9<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u53ef\u6269\u5c55\u7684\u540e\u8bad\u7ec3 RL \u5e93<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u652f\u6301 GRPO\u3001SFT\u3001DPO\u3001DAPO \u7b97\u6cd5<\/li><li>Ray \u8d44\u6e90\u7ba1\u7406<\/li><li>Megatron Core \u5e76\u884c<\/li><li>\u8bad\u7ec3\u4e86 Nemotron-3-Nano-30B<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: NVIDIA \u751f\u6001\u7684 RL \u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"47-nemo-gym--\">4.7 NeMo Gym \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/NVIDIA-NeMo\/Gym\">https:\/\/github.com\/NVIDIA-NeMo\/Gym<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: NVIDIA \u5b98\u65b9<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u4e3a LLM \u8bad\u7ec3\u6784\u5efa RL \u73af\u5883<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u591a\u6b65\u9aa4\u3001\u591a\u8f6e\u73af\u5883\u811a\u624b\u67b6<\/li><li>\u4e0e NeMo RL\u3001OpenRLHF\u3001TRL\u3001Unsloth \u4e92\u64cd\u4f5c<\/li><li>\u5185\u7f6e Reasoning Gym \u548c Aviary \u73af\u5883<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u6784\u5efa\u81ea\u5b9a\u4e49 RL \u8bad\u7ec3\u73af\u5883<\/li><\/ul>\n\n\n\n<h3 id=\"48-rllm-\">4.8 rLLM \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/rllm-org\/rllm\">https:\/\/github.com\/rllm-org\/rllm<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u540e\u8bad\u7ec3\uff08Post-training\uff09RL \u6846\u67b6\uff0c\u4e13\u6ce8\u4e8e Agent \u8bad\u7ec3<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u81ea\u5b9a\u4e49 Agent + \u73af\u5883 \u2192 RL \u8bad\u7ec3 \u2192 \u90e8\u7f72<\/li><li>v0.2: AgentWorkflowEngine\uff0c\u652f\u6301\u4efb\u610f Agentic \u7a0b\u5e8f\u8bad\u7ec3<\/li><li>LoRA + VLM \u8bad\u7ec3\u652f\u6301<\/li><li><strong>\u5b9e\u9645\u6848\u4f8b<\/strong>: rLLM-FinQA-4B\uff084B \u53c2\u6570\uff09\u5728\u91d1\u878d\u5206\u6790\u4efb\u52a1\u4e0a\u8d85\u8d8a Qwen3-235B<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: Agent \u80fd\u529b RL \u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"49-ragen--\">4.9 RAGEN \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/RAGEN-AI\/RAGEN\">https:\/\/github.com\/RAGEN-AI\/RAGEN<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u591a\u8f6e RL \u6846\u67b6\uff0c\u4e13\u6ce8\u63a8\u7406 Agent \u8bad\u7ec3<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>StarPO \u6846\u67b6<\/li><li>10 \u4e2a\u5185\u7f6e\u73af\u5883\uff08Sokoban\u3001WebShop\u3001Lean \u7b49\uff09<\/li><li>\u53d1\u73b0\u5e76\u5b9a\u4e49\u4e86 &#8220;Echo Trap&#8221; \u8bad\u7ec3\u4e0d\u7a33\u5b9a\u6027\u95ee\u9898<\/li><li>RAGEN V2\uff082026-03\uff09<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u63a8\u7406 Agent \u7684\u591a\u8f6e RL \u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"410-f-grpo--\">4.10 f-GRPO \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/rhaldarpurdue\/f-GRPO\">https:\/\/github.com\/rhaldarpurdue\/f-GRPO<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2602.05946\">https:\/\/arxiv.org\/abs\/2602.05946<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u57fa\u4e8e f-\u6563\u5ea6\u7684 GRPO \u6269\u5c55\uff0c\u63d0\u4f9b\u66f4\u7075\u6d3b\u7684\u7b56\u7565\u4f18\u5316<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u652f\u6301 KL\u3001\u53cd\u5411 KL\u3001Pearson\u3001Hellinger\u3001Jensen-Shannon\u3001\u5168\u53d8\u5dee\u7b49\u591a\u79cd f-\u6563\u5ea6<\/li><li>\u5728 RLVR\uff08\u6570\u5b66\u63a8\u7406\uff09\u548c PA\uff08\u5b89\u5168\u5bf9\u9f50\uff09\u4efb\u52a1\u4e0a\u5747\u4f18\u4e8e\u6807\u51c6 GRPO<\/li><li>\u57fa\u4e8e Unsloth \u6846\u67b6\u6784\u5efa<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u66f4\u7cbe\u7ec6\u63a7\u5236\u7b56\u7565\u4f18\u5316\u65b9\u5411\u7684 RL \u5bf9\u9f50\u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"411-tree-grpo--\">4.11 Tree-GRPO \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/AMAP-ML\/Tree-GRPO\">https:\/\/github.com\/AMAP-ML\/Tree-GRPO<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: ICLR 2026<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u7528\u6811\u641c\u7d22\u66ff\u4ee3\u94fe\u5f0f\u91c7\u6837\uff0c\u5927\u5e45\u964d\u4f4e RL \u8bad\u7ec3\u7684\u91c7\u6837\u5f00\u9500<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u57fa\u4e8e ReAct \u6b65\u9aa4\u7ea7\u8282\u70b9\u7684\u8bed\u4e49\u641c\u7d22\u6811<\/li><li>\u5171\u4eab\u524d\u7f00\u4f7f\u5f97\u76f8\u540c token \u9884\u7b97\u4e0b\u53ef\u83b7 <strong>4 \u500d<\/strong> \u91c7\u6837\u91cf<\/li><li>\u6811\u7ed3\u6784\u81ea\u7136\u751f\u6210\u6b65\u9aa4\u7ea7\u8fc7\u7a0b\u76d1\u7763\u4fe1\u53f7\uff08\u4ec5\u9700\u7ed3\u679c\u5956\u52b1\uff09<\/li><li>\u57fa\u4e8e Qwen2.5-3B \u9a8c\u8bc1<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: LLM Agent \u7684 RL \u8bad\u7ec3\uff0c\u5c24\u5176\u9002\u5408\u591a\u6b65\u63a8\u7406\u548c\u5de5\u5177\u8c03\u7528<\/li><\/ul>\n\n\n\n<h3 id=\"412-simplerl-reason--\">4.12 SimpleRL-Reason \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/hkust-nlp\/simpleRL-reason\">https:\/\/github.com\/hkust-nlp\/simpleRL-reason<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u9999\u6e2f\u79d1\u6280\u5927\u5b66<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u6781\u7b80\u63a8\u7406 RL \u8bad\u7ec3\u65b9\u6848<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>DeepSeek-R1 \u98ce\u683c\u8bad\u7ec3<\/li><li>7B \u6a21\u578b\u4ec5\u9700 8K \u6837\u672c\u5373\u8fbe 33.3% AIME\u300177.2% MATH<\/li><li>\u57fa\u4e8e\u89c4\u5219\u7684\u5956\u52b1\uff0c\u65e0\u9700 SFT \u6216 Reward Model<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4f4e\u8d44\u6e90\u63a8\u7406\u80fd\u529b\u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"413-swe-rl--\">4.13 SWE-RL \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/facebookresearch\/swe-rl\">https:\/\/github.com\/facebookresearch\/swe-rl<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: Meta\uff08NeurIPS 2025\uff09<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u9762\u5411\u8f6f\u4ef6\u5de5\u7a0b\u63a8\u7406\u7684 RL \u8bad\u7ec3<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>Llama3-SWE-RL-70B \u5728 SWE-bench Verified \u4e0a\u8fbe 41%<\/li><li>\u4ece\u5f00\u6e90\u8f6f\u4ef6\u6f14\u8fdb\u6570\u636e\u4e2d\u5b66\u4e60<\/li><li>Agentless Mini \u81ea\u52a8\u4ee3\u7801\u4fee\u590d\u6846\u67b6<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u8bad\u7ec3\u7f16\u7a0b Agent<\/li><\/ul>\n\n\n\n<h3 id=\"414-openmanus-rl--\">4.14 OpenManus-RL \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/OpenManus\/OpenManus-RL\">https:\/\/github.com\/OpenManus\/OpenManus-RL<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: UIUC + MetaGPT \u56e2\u961f<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u9762\u5411 LLM Agent \u7684 RL \u5fae\u8c03<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>PPO + AgentGym \u73af\u5883 + verl \u8bad\u7ec3<\/li><li>\u5728 GAIA\u3001AgentBench\u3001WebShop\u3001OSWorld \u4e0a\u6d4b\u8bd5<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u901a\u7528 Agent \u80fd\u529b RL \u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"415-reasoning-gym--\">4.15 Reasoning Gym \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/open-thought\/reasoning-gym\">https:\/\/github.com\/open-thought\/reasoning-gym<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: NeurIPS 2025 Spotlight<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u7a0b\u5e8f\u5316\u63a8\u7406\u73af\u5883\u96c6\u5408<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>100+ \u4efb\u52a1\u8986\u76d6\u4ee3\u6570\u3001\u7b97\u6cd5\u3001\u903b\u8f91\u3001\u6e38\u620f\u3001\u51e0\u4f55<\/li><li>\u65e0\u9650\u53ef\u63a7\u4efb\u52a1\u751f\u6210<\/li><li>\u5df2\u9a8c\u8bc1\u8de8\u9886\u57df\u8fc1\u79fb\u6548\u679c<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: RLVR\uff08\u57fa\u4e8e\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684 RL\uff09\u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"416-llamagym-\">4.16 LlamaGym \u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/KhoomeiK\/LlamaGym\">https:\/\/github.com\/KhoomeiK\/LlamaGym<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u7528\u5728\u7ebf\u5f3a\u5316\u5b66\u4e60\u5fae\u8c03 LLM Agent<\/li><li><strong>\u7279\u70b9<\/strong>: \u5b9a\u4e49 Agent \u2192 \u521b\u5efa LLM \u2192 \u7f16\u5199 RL Loop<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5feb\u901f\u5b9e\u9a8c Agent RL \u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"417-art-agent-reinforcement-trainer--\">4.17 ART (Agent Reinforcement Trainer) \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/OpenPipe\/ART\">https:\/\/github.com\/OpenPipe\/ART<\/a><\/li><li><strong>\u5b98\u65b9\u6587\u6863<\/strong>: <a href=\"https:\/\/art.openpipe.ai\">https:\/\/art.openpipe.ai<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: OpenPipe<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u5f00\u6e90 RL \u6846\u67b6\uff0c\u4e13\u4e3a<strong>\u591a\u6b65\u9aa4\u771f\u5b9e\u4efb\u52a1 Agent \u8bad\u7ec3<\/strong>\u8bbe\u8ba1\u2014\u2014\u628a GRPO \u5f15\u5165\u4efb\u610f Python Agent \u5e94\u7528<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u539f\u751f\u591a\u8f6e + \u5de5\u5177\u8c03\u7528\u652f\u6301<\/strong>\uff1a\u5c06\u6bcf\u6b21 Agent \u8fd0\u884c\u8bb0\u5f55\u4e3a\u5b8c\u6574 Trajectory\uff08\u884c\u52a8\u5386\u53f2\uff09\uff0c\u5929\u7136\u9002\u914d\u5de5\u5177\u8c03\u7528\u548c\u591a\u8f6e\u5bf9\u8bdd<\/li><li>vLLM \u8d1f\u8d23\u63a8\u7406 + Unsloth-powered GRPO \u8d1f\u8d23\u8bad\u7ec3\uff1b\u6bcf\u8f6e\u8bad\u7ec3\u540e\u81ea\u52a8\u52a0\u8f7d\u65b0 LoRA checkpoint \u5230\u63a8\u7406\u670d\u52a1\u5668<\/li><li><strong>LangGraph \u96c6\u6210<\/strong>\uff1a\u76f4\u63a5\u5bf9 LangGraph Agent \u8fdb\u884c RL \u8bad\u7ec3<\/li><li><strong>MCP\u00b7RL<\/strong>\uff1a\u81ea\u52a8\u8bad\u7ec3\u6a21\u578b\u9ad8\u6548\u4f7f\u7528\u4efb\u610f MCP Server \u5de5\u5177<\/li><li>RULER\uff1a\u81ea\u52a8\u5956\u52b1\u751f\u6210\uff0c\u7b80\u5316 reward function \u8bbe\u8ba1<\/li><li>\u5ba2\u6237\u7aef\u53ef\u5728\u672c\u5730\u8fd0\u884c\uff0c\u670d\u52a1\u7aef\u542f\u52a8\u4e34\u65f6 GPU \u73af\u5883\uff0c\u89e3\u8026\u90e8\u7f72<\/li><li>\u652f\u6301 Qwen3.5\u3001Llama\u3001GPT-OSS \u7b49\u4e3b\u6d41\u6a21\u578b<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u628a\u73b0\u6709\u591a\u6b65\u9aa4 Agent \u5e94\u7528\uff08LangGraph \/ MCP\uff09\u5347\u7ea7\u4e3a\u53ef\u901a\u8fc7 RL \u81ea\u6211\u4f18\u5316\u7684\u667a\u80fd Agent<\/li><\/ul>\n\n\n\n<h3 id=\"418-marti--\">4.18 MARTI \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/TsinghuaC3I\/MARTI\">https:\/\/github.com\/TsinghuaC3I\/MARTI<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u6e05\u534e\u5927\u5b66 C3I\uff08\u667a\u80fd\u8ba1\u7b97\uff09\u7814\u7a76\u7ec4<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u9762\u5411 LLM \u591a Agent \u7cfb\u7edf\u7684\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u4e0e\u63a8\u7406\u6846\u67b6\u2014\u2014\u5c06\u4e2d\u5fc3\u5316\u591a Agent \u4ea4\u4e92\u4e0e\u5206\u5e03\u5f0f\u7b56\u7565\u8bad\u7ec3\u7edf\u4e00\u96c6\u6210<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u57fa\u4e8e OpenRLHF \u6784\u5efa\uff0c\u6574\u5408\u6700\u65b0 RL \u8bad\u7ec3\u6280\u672f\uff0c\u652f\u6301\u5f02\u8d28\u591a Agent \u8bad\u7ec3<\/li><li><strong>\u591a\u8f6e\u5f02\u6b65 Rollout<\/strong>\uff1a\u63d0\u5347\u8bad\u7ec3\u6548\u7387\uff0c\u652f\u6301\u52a8\u6001\u5de5\u4f5c\u6d41<\/li><li>\u52a8\u6001\u5956\u52b1\u7cfb\u7edf\uff1a\u540c\u65f6\u652f\u6301\u57fa\u4e8e\u89c4\u5219\u7684\u53ef\u9a8c\u8bc1\u5956\u52b1 + LLM \u751f\u6210\u5956\u52b1<\/li><li>\u9ad8\u7ea7 RL \u6280\u672f\uff1aGSPO loss\uff08\u5e8f\u5217\u7ea7\u4f18\u5316\uff09\u3001TIS \u7ea0\u6b63\u3001\u52a8\u6001\u6570\u636e\u8fc7\u6ee4\u3001\u8d85\u957f\u5e8f\u5217\u7f13\u51b2\uff08\u6700\u9ad8 <strong>32K token<\/strong>\uff09<\/li><li><strong>MARTI-v2\uff082026-02\uff09<\/strong>\uff1aMARS2\u2014\u2014\u6811\u641c\u7d22\u589e\u5f3a\u7684 RL\uff0c\u4e13\u4e3a\u590d\u6742\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8bbe\u8ba1\uff0c\u81ea\u9002\u5e94\u8282\u70b9\u6269\u5c55\u4e0e\u7cbe\u70bc<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u591a Agent \u7cfb\u7edf\u7684 RL \u8bad\u7ec3\uff1b\u8d85\u957f\u5e8f\u5217 RL\uff1b\u4ee3\u7801\u751f\u6210 Agent \u6811\u641c\u7d22 RL<\/li><\/ul>\n\n\n\n<h3 id=\"419-arctic-rlsnowflake-\">4.19 Arctic RL\uff08Snowflake\uff09\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Snowflake-AI-Research\/Arctic-Platform\">https:\/\/github.com\/Snowflake-AI-Research\/Arctic-Platform<\/a><\/li><li><strong>\u5b98\u65b9\u535a\u5ba2<\/strong>: <a href=\"https:\/\/www.snowflake.com\/en\/blog\/engineering\/arctic-rl-open-source-backend\/\">https:\/\/www.snowflake.com\/en\/blog\/engineering\/arctic-rl-open-source-backend\/<\/a><\/li><li><strong>\u53d1\u5e03<\/strong>: 2026-06-29<\/li><li><strong>\u6765\u6e90<\/strong>: Snowflake AI Research<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u9762\u5411\u4f01\u4e1a\u7ea7 RL \u540e\u8bad\u7ec3\u7684\u5f00\u6e90\u7cfb\u7edf\u57fa\u7840\u8bbe\u65bd\u4e0e\u540e\u7aef\u5c42\u2014\u2014\u628a RL \u7b97\u6cd5\u903b\u8f91\u4e0e\u5e95\u5c42 GPU \u7cfb\u7edf\u4f18\u5316\u89e3\u8026\uff0c\u8ba9\u6846\u67b6\u5f00\u53d1\u8005\u4e13\u6ce8\u7b97\u6cd5<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>ZoRRo\uff08Zero Redundancy Rollouts\uff0c\u96f6\u5197\u4f59 Rollout\uff09<\/strong>\uff1aRL \u8bad\u7ec3\uff08PPO\/GRPO\uff09\u4e2d\u540c\u4e00 prompt \u88ab\u53cd\u590d\u91c7\u6837\uff0c\u957f\u5e8f\u5217\u4efb\u52a1\u4e0b 80\u201395% \u7684 token \u662f\u91cd\u590d\u7684 prompt token\uff0c\u914d\u5408\u6ce8\u610f\u529b O(n\u00b2) \u5f00\u9500\u6210\u4e3a\u4e3b\u8981\u6210\u672c\uff1bZoRRo \u81ea\u52a8\u53bb\u91cd\u3001\u6bcf\u4e2a\u552f\u4e00 prompt \u53ea\u6253\u5305\/\u524d\u5411\u4e00\u6b21<\/li><li><strong>\u6027\u80fd<\/strong>\uff1aActor \u66f4\u65b0\u6700\u9ad8 <strong>6x \u52a0\u901f<\/strong>\u3001\u7aef\u5230\u7aef <strong>3.5x \u52a0\u901f<\/strong>\uff1bArctic-Text2SQL-R2 \u8bad\u7ec3\u4ece\u7ea6 5 \u5929\u7f29\u77ed\u5230 32 \u5f20 H200 \u4e0a\u7ea6 36 \u5c0f\u65f6<\/li><li>\u4f5c\u4e3a\u9ad8\u541e\u5410 RL \u8bad\u7ec3\/\u63a8\u7406\u540e\u7aef\uff0c\u53ef<strong>\u63d2\u5165\u73b0\u6709 RL \u6846\u67b6<\/strong>\uff0c\u975e\u66ff\u4ee3\u5f0f\u91cd\u5199<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u957f\u4e0a\u4e0b\u6587 RL \u8bad\u7ec3\u7684\u7cfb\u7edf\u7ea7\u63d0\u901f\uff1b\u4f01\u4e1a\u7ea7\u5927\u89c4\u6a21 RLHF\/RLVR \u964d\u672c<\/li><\/ul>\n\n\n\n<h3 id=\"420-openrlgoogle-gke-labs-\">4.20 OpenRL\uff08Google GKE Labs\uff09\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/gke-labs\/open-rl\">https:\/\/github.com\/gke-labs\/open-rl<\/a><\/li><li><strong>\u5b98\u65b9\u535a\u5ba2<\/strong>: <a href=\"https:\/\/opensource.googleblog.com\/2026\/06\/introducing-openrl-a-self-hosted-post-training-api-for-fine-tuning-llms.html\">https:\/\/opensource.googleblog.com\/2026\/06\/introducing-openrl-a-self-hosted-post-training-api-for-fine-tuning-llms.html<\/a><\/li><li><strong>\u53d1\u5e03<\/strong>: 2026-06<\/li><li><strong>\u6765\u6e90<\/strong>: Google GKE Labs<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u81ea\u6258\u7ba1\u7684\u540e\u8bad\u7ec3\/\u5fae\u8c03 API\u2014\u2014\u628a RL \u57fa\u7840\u8bbe\u65bd\u4ece AI \u7814\u7a76\u4e2d\u62bd\u8c61\u51fa\u6765\uff0c\u8ba9 ML \u56e2\u961f\u5728\u81ea\u5df1\u7684 Kubernetes \u96c6\u7fa4\u4e0a\u6269\u5c55\u540e\u8bad\u7ec3\u5de5\u4f5c\u6d41<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u5b9e\u73b0 <strong>Tinker \u517c\u5bb9 API<\/strong>\uff1a\u7528 Tinker SDK \u4ece\u672c\u5730\u673a\u5668\u7528\u547d\u4ee4\u5f0f Python \u4ee3\u7801\u7f16\u6392 RL \u8bad\u7ec3\u5faa\u73af<\/li><li>\u53ef\u5728\u5355\u673a\u6216 Kubernetes \u96c6\u7fa4\u4e0a\u8fd0\u884c\uff0c\u8d77\u6b65\u805a\u7126 LoRA \u5fae\u8c03<\/li><li><strong>\u591a RL \u4efb\u52a1\u5e76\u884c\u8c03\u5ea6<\/strong>\u63d0\u5347\u6574\u4f53 GPU \u5229\u7528\u7387\u2014\u2014\u4f20\u7edf RL \u5faa\u73af\u4e25\u683c\u4e32\u884c\uff0creward \u8ba1\u7b97\u7b49 CPU\/\u7f51\u7edc\u4efb\u52a1\u5e38\u4f7f GPU \u7a7a\u95f2<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u5728\u81ea\u6709\u57fa\u7840\u8bbe\u65bd\uff08K8s\uff09\u4e0a\u81ea\u6258\u7ba1\u3001\u53ef\u6269\u5c55\u540e\u8bad\u7ec3\u7684\u56e2\u961f\uff1bTinker \u751f\u6001\u7528\u6237<\/li><\/ul>\n\n\n\n<h3 id=\"421-openclaw-rl--\">4.21 OpenClaw-RL \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Gen-Verse\/OpenClaw-RL\">https:\/\/github.com\/Gen-Verse\/OpenClaw-RL<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/huggingface.co\/papers\/2603.10165\">https:\/\/huggingface.co\/papers\/2603.10165<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u666e\u6797\u65af\u987f\u5927\u5b66 \/ Gen-Verse<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: &#8220;Train any agent simply by talking&#8221;\u2014\u2014\u5168\u5f02\u6b65 RL \u6846\u67b6\uff0c\u628a\u65e5\u5e38\u5bf9\u8bdd\u53d8\u6210\u8bad\u7ec3\u4fe1\u53f7\uff0c\u5728<strong>\u5728\u7ebf\u90e8\u7f72\u671f\u95f4\u6301\u7eed\u4f18\u5316<\/strong> Agent \u7b56\u7565\uff0c\u4e0d\u6253\u65ad\u4f7f\u7528<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u628a\u81ea\u6258\u7ba1\u6a21\u578b\u5305\u88c5\u4e3a OpenAI \u517c\u5bb9 API\uff0c\u62e6\u622a\u5b9e\u65f6\u591a\u8f6e\u5bf9\u8bdd\u5e76\u5728\u540e\u53f0\u6301\u7eed\u4f18\u5316\u7b56\u7565<\/li><li><strong>\u56db\u7ec4\u4ef6\u5168\u5f02\u6b65\u89e3\u8026<\/strong>\uff1a\u7b56\u7565\u670d\u52a1\u3001\u73af\u5883\u6267\u884c\u3001\u5956\u52b1\u8bc4\u5224\u3001\u6a21\u578b\u8bad\u7ec3\u5404\u81ea\u72ec\u7acb\u5f02\u6b65\u5faa\u73af<\/li><li>\u6355\u83b7\u8f6c\u77ac\u5373\u901d\u7684&#8221;\u4e0b\u4e00\u72b6\u6001\u4fe1\u53f7&#8221;\uff08\u5982\u7528\u6237\u7ea0\u6b63\u3001\u7ec8\u7aef trace \u62a5\u9519\uff09\uff0c\u7528\u6807\u91cf\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff08PRM\uff09+ token \u7ea7 Hindsight-Guided On-Policy Distillation\uff08OPD\uff09\u8f6c\u6362\u4e3a\u4f18\u5316\u68af\u5ea6<\/li><li><strong>\u96f6\u4eba\u5de5\u6807\u6ce8<\/strong>\uff1b\u4e09\u79cd\u5b66\u4e60\u8303\u5f0f\uff08Binary RL \/ OPD \/ Combine\uff09\uff1b\u540c\u65f6\u652f\u6301\u4e2a\u4eba Agent \u4e0e\u901a\u7528 Agent\uff1b\u5927\u89c4\u6a21\u73af\u5883\u5e76\u884c<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u90e8\u7f72\u5373\u8bad\u7ec3\u7684\u5e38\u9a7b\u5f0f\u5728\u7ebf RL\uff1b\u4e2a\u6027\u5316 Agent \u6301\u7eed\u81ea\u6211\u6539\u8fdb<\/li><\/ul>\n\n\n\n<h3 id=\"422-vimevllm-\u5b98\u65b9-\">4.22 vime\uff08vLLM \u5b98\u65b9\uff09\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/vllm-project\/vime\">https:\/\/github.com\/vllm-project\/vime<\/a><\/li><li><strong>\u5b98\u65b9\u535a\u5ba2<\/strong>: <a href=\"https:\/\/vllm.ai\/blog\/2026-06-09-announcing-vime\">https:\/\/vllm.ai\/blog\/2026-06-09-announcing-vime<\/a><\/li><li><strong>\u53d1\u5e03<\/strong>: 2026-06-09<\/li><li><strong>\u6765\u6e90<\/strong>: vLLM \u793e\u533a\uff08\u5b98\u65b9\u9879\u76ee\uff09<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: vLLM \u751f\u6001\u539f\u751f\u7684 RL \u540e\u8bad\u7ec3\u6846\u67b6\u2014\u2014&#8221;\u7b80\u5355\u3001\u7a33\u5b9a\u3001\u9ad8\u6548&#8221;\uff0c\u628a Megatron \u8bad\u7ec3\u4e0e vLLM \u63a8\u7406\u7edf\u4e00\u5230\u4e00\u6761\u6d41\u6c34\u7ebf<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u6cbf\u7528 slime \u6210\u719f\u7684\u8bad\u7ec3\u6808\u4e0e\u6570\u636e\u751f\u6210\u8bbe\u8ba1\uff0c\u9ed8\u8ba4\u7528 <strong>vLLM\uff08vllm-router\uff09\u4f5c\u4e3a rollout \u540e\u7aef<\/strong><\/li><li><strong>\u4e09\u9636\u6bb5\u89e3\u8026\u7684\u8bad\u63a8\u67b6\u6784<\/strong>\uff1a\u8bad\u7ec3\uff08Megatron\uff09\u8d1f\u8d23\u53c2\u6570\u66f4\u65b0\u4e0e\u6743\u91cd\u540c\u6b65\uff0crollout\uff08vLLM + Router\uff09\u8d1f\u8d23\u63a8\u7406\u91c7\u6837\u5e76\u4ea7\u51fa\u5e26\u5956\u52b1\/\u9a8c\u8bc1\u4fe1\u53f7\u7684\u8bad\u7ec3\u6837\u672c<\/li><li>\u5168\u5f02\u6b65\u6d41\u6c34\u7ebf + <strong>\u8bad\u63a8\u4e0d\u4e00\u81f4\uff08train-inference mismatch\uff09\u7ea0\u6b63<\/strong>\uff0c\u4fdd\u8bc1\u8bad\u7ec3\u7a33\u5b9a<\/li><li>\u539f\u751f\u652f\u6301 <strong>Agentic RL<\/strong>\uff1a\u591a\u8f6e\u5de5\u5177\u8c03\u7528\u4e0e\u591a Agent \u573a\u666f\uff1b\u5feb\u901f\u8ddf\u8fdb MoE\u3001VLM \u7b49\u65b0\u67b6\u6784<\/li><li>\u7ee7\u627f slime \u7684\u5e7f\u6cdb\u6a21\u578b\u652f\u6301\uff08Qwen \u7cfb\u5217\u3001DeepSeek V3 \u7b49\uff09\uff1b\u5df2\u83b7 <strong>AMD ROCm<\/strong> \u539f\u751f\u652f\u6301\uff08MI355X \u7aef\u5230\u7aef RL\uff09<\/li><li>Apache-2.0 \u8bb8\u53ef<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5df2\u5728\u7528 vLLM \u63a8\u7406\u6808\u3001\u5e0c\u671b\u7528\u540c\u4e00\u751f\u6001\u505a RL \u540e\u8bad\u7ec3\u7684\u56e2\u961f\uff1b\u8ffd\u6c42\u8bad\u63a8\u5bf9\u9f50\u7a33\u5b9a\u6027\u7684 RL \u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"423-dllm-rltracerl-\">4.23 dLLM-RL\uff08TraceRL\uff09\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Gen-Verse\/dLLM-RL\">https:\/\/github.com\/Gen-Verse\/dLLM-RL<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2509.06949\uff08TraceRL\">https:\/\/arxiv.org\/abs\/2509.06949\uff08TraceRL<\/a>\uff0cICLR 2026\uff09<\/li><li><strong>\u6765\u6e90<\/strong>: \u666e\u6797\u65af\u987f \/ Gen-Verse\uff08\u6768\u51cc\u7b49\uff09<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u9996\u4e2a\u9762\u5411<strong>\u6269\u6563\u8bed\u8a00\u6a21\u578b\uff08Diffusion LLM\uff09<\/strong>\u7684\u5f3a\u5316\u5b66\u4e60\u540e\u8bad\u7ec3\u6846\u67b6\u2014\u2014\u628a\u300c\u504f\u597d\u63a8\u7406\u8f68\u8ff9\u300d\u5f15\u5165 diffusion LLM \u7684\u540e\u8bad\u7ec3<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>TraceRL\uff08\u8f68\u8ff9\u611f\u77e5 RL\uff09<\/strong>\uff1a\u76f4\u63a5\u4f18\u5316\u504f\u597d\u63a8\u7406\u8f68\u8ff9\uff0c\u8de8\u4e0d\u540c\u6269\u6563\u67b6\u6784\u901a\u7528<\/li><li>\u5f15\u5165<strong>\u57fa\u4e8e\u6269\u6563\u7684\u4ef7\u503c\u6a21\u578b<\/strong>\u589e\u5f3a\u8bad\u7ec3\u7a33\u5b9a\u6027\u3001\u964d\u4f4e\u65b9\u5dee\uff1b\u53ef\u9009\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\u505a\u7ec6\u7c92\u5ea6\u76d1\u7763<\/li><li>\u5b8c\u6574\u8986\u76d6 diffusion LLM \u7684 deployment \/ SFT \/ RL \/ RLHF\uff0c\u6a2a\u8de8\u6570\u5b66\u3001\u4ee3\u7801\u3001\u591a\u6a21\u6001\u591a\u79cd\u8bbe\u5b9a<\/li><li>\u9a71\u52a8 <strong>SOTA TraDo \u7cfb\u5217<\/strong>\uff1aTraDo-8B-Instruct \u5728\u6570\u5b66\u63a8\u7406\u4e0a\u76f8\u5bf9 Qwen2.5-7B-Instruct \u63d0\u5347 6.1%\uff1b\u901a\u8fc7\u8bfe\u7a0b\u5b66\u4e60\u5f97\u5230\u9996\u4e2a long-CoT \u6269\u6563 LLM<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u6269\u6563\u8bed\u8a00\u6a21\u578b\u7684 RL \u540e\u8bad\u7ec3\uff1b\u4e0e Open-dLLM \u7b49\u6269\u6563 LLM \u5168\u6808\u914d\u5408\u4f7f\u7528<\/li><\/ul>\n\n\n\n<h3 id=\"424-mileslmsys--sglang-\u56e2\u961f-\">4.24 Miles\uff08LMSYS \/ SGLang \u56e2\u961f\uff09\u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/radixark\/miles\">https:\/\/github.com\/radixark\/miles<\/a><\/li><li><strong>\u5b98\u65b9\u535a\u5ba2<\/strong>: <a href=\"https:\/\/www.lmsys.org\/blog\/2026-08-18-miles-v0-1\/\">https:\/\/www.lmsys.org\/blog\/2026-08-18-miles-v0-1\/<\/a><\/li><li><strong>\u6587\u6863<\/strong>: <a href=\"https:\/\/radixark.mintlify.app\/\">https:\/\/radixark.mintlify.app\/<\/a><\/li><li><strong>\u53d1\u5e03<\/strong>: v0.1 \u2014 2026-08-18<\/li><li><strong>\u6765\u6e90<\/strong>: LMSYS Org \/ SGLang \u56e2\u961f\uff08RadixArk \u7ef4\u62a4\uff09<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u9762\u5411<strong>\u4f01\u4e1a\u751f\u4ea7\u7ea7<\/strong>\u7684\u5927\u89c4\u6a21 RL \u540e\u8bad\u7ec3\u6846\u67b6\u2014\u2014\u4ece slime fork \u5e76\u4e0e\u4e4b\u534f\u540c\u6f14\u8fdb\uff0c\u4e13\u6ce8\u5927\u89c4\u6a21 MoE \u8bad\u7ec3\u4e0e\u751f\u4ea7\u843d\u5730<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>SGLang \u9ad8\u541e\u5410 rollout + Megatron-LM \u53ef\u6269\u5c55\u8bad\u7ec3<\/strong>\uff1a\u5728\u4e07\u4ebf\u53c2\u6570\u89c4\u6a21\u4e0b\u63d0\u4f9b RL \u8fd0\u884c\u771f\u6b63\u9700\u8981\u7684\u7cbe\u5ea6\u3001\u7a33\u5b9a\u6027\u4e0e\u53ef\u89c2\u6d4b\u6027\u80fd\u529b<\/li><li><strong>\u5168\u5f02\u6b65 RL<\/strong>\uff1arollout worker \u4e0e training worker \u89e3\u8026\uff0con-policy \/ off-policy \u8c03\u5ea6\u53ef\u914d\u7f6e\uff1b\u6d41\u6c34\u7ebf\u9488\u5bf9\u51cf\u5c11\u6c14\u6ce1\u8c03\u4f18\uff0c\u5f02\u6b65 rollout \u4e0e eval \u6a21\u5f0f\u53ef\u81ea\u5b9a\u4e49<\/li><li><strong>\u5b8c\u6574 Agentic \u8bad\u7ec3\u5de5\u4f5c\u6d41<\/strong>\uff1a\u591a\u8f6e\u4f1a\u8bdd\u3001\u5de5\u5177\u6267\u884c\u3001\u6c99\u7bb1\u73af\u5883\u3001token \u7ea7\u5fe0\u5b9e\u8f68\u8ff9\u6355\u83b7<\/li><li><strong>LoRA \/ multi-LoRA<\/strong>\uff1a\u7528\u4e00\u5c0f\u90e8\u5206 GPU \u8bad\u7ec3\u524d\u6cbf\u89c4\u6a21\u6a21\u578b\uff0c\u540c\u4e00\u6279 adapter \u53ef\u76f4\u63a5\u52a0\u8f7d\u8fdb SGLang \u505a rollout<\/li><li>SGLang \u7684 memory-aware sleep\/wake \u673a\u5236\u5728 RL \u6bcf\u6b65\u4e4b\u95f4\u91ca\u653e\u5e76\u6062\u590d KV cache \u4e0e\u6743\u91cd\uff0c\u907f\u514d\u91cd\u590d\u78c1\u76d8 I\/O \u4e0e CUDA graph \u91cd\u6355\u83b7<\/li><li>\u5df2\u83b7 <strong>AMD Instinct\uff08ROCm\uff09<\/strong> \u652f\u6301<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4f01\u4e1a\u7ea7\u5927\u89c4\u6a21 MoE \u6a21\u578b RL \u540e\u8bad\u7ec3\uff1b\u5df2\u4f7f\u7528 SGLang \u63a8\u7406\u6808\u3001\u8ffd\u6c42\u751f\u4ea7\u7a33\u5b9a\u6027\u4e0e\u53ef\u89c2\u6d4b\u6027\u7684\u56e2\u961f<\/li><\/ul>\n\n\n\n<h3 id=\"425-open-agentrlgen-verse-\">4.25 Open-AgentRL\uff08Gen-Verse\uff09\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Gen-Verse\/Open-AgentRL\">https:\/\/github.com\/Gen-Verse\/Open-AgentRL<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: Gen-Verse\uff08\u666e\u6797\u65af\u987f\u7b49\uff0cRLAnything \u4e0e AutoTool \u5747\u4e3a ICML 2026\uff09<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u9762\u5411 LLM \u4e0e Agentic \u573a\u666f\u7684\u5f00\u6e90 RL \u5168\u5bb6\u6876\u2014\u2014\u4e00\u5957\u7b97\u6cd5\u901a\u5403\u7ec8\u7aef\u3001GUI\u3001\u8f6f\u4ef6\u5de5\u7a0b\u3001\u5de5\u5177\u8c03\u7528\u7b49\u5f02\u6784 Agent \u73af\u5883<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>RLAnything\uff08ICML 2026\uff09<\/strong>\uff1a\u901a\u7528\u4e14\u53ef\u6269\u5c55\u7684 Agentic RL \u7b97\u6cd5\uff0c\u8de8 terminal \/ GUI \/ SWE \/ tool-call \u8bbe\u5b9a\u901a\u7528\uff1bQwen3-VL-8B-Thinking \u5728 OSWorld \u4e0a +9.1%\uff0cQwen2.5-7B-Instruct \u5728 AlfWorld \u4e0a +18.7%\u3001LiveBench \u4e0a +11.9%<\/li><li><strong>AutoTool\uff08ICML 2026\uff09<\/strong>\uff1a\u8ba9 Agent \u5728\u63a8\u7406\u8f68\u8ff9\u4e2d\u5177\u5907<strong>\u52a8\u6001\u5de5\u5177\u9009\u62e9<\/strong>\u80fd\u529b\uff0c\u800c\u975e\u56fa\u5b9a\u5de5\u5177\u96c6<\/li><li><strong>DemyAgent<\/strong>\uff1a\u4ec5 4B \u53c2\u6570\u7684 DemyAgent-4B \u5728\u591a\u4e2a\u9ad8\u96be\u57fa\u51c6\u4e0a\u5339\u654c\u751a\u81f3\u8d85\u8d8a 14B\/32B \u6a21\u578b\uff0c\u8d85\u8fc7 ReTool-32B\u3001rStar2-Agent-14B\uff0c\u8fbe\u5230 SOTA Agentic \u63a8\u7406\u6c34\u5e73<\/li><li>\u662f <strong>OpenClaw-RL\uff084.21\uff09<\/strong> \u6240\u57fa\u4e8e\u7684\u5e95\u5c42\u6846\u67b6<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u7528\u540c\u4e00\u5957 RL \u7b97\u6cd5\u8986\u76d6\u591a\u7c7b Agent \u73af\u5883\uff1b\u5c0f\u53c2\u6570\u91cf\u9ad8\u6027\u80fd Agent \u6a21\u578b\u8bad\u7ec3<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"5-\u81ea\u52a8\u5316\u8d85\u53c2\u6570\u4f18\u5316--automl-agent\">5. \u81ea\u52a8\u5316\u8d85\u53c2\u6570\u4f18\u5316 \/ AutoML Agent<\/h2>\n\n\n\n<h3 id=\"51-agenthpo-\">5.1 AgentHPO \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2402.01881\">https:\/\/arxiv.org\/abs\/2402.01881<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: LLM \u9a71\u52a8\u8d85\u53c2\u6570\u4f18\u5316<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u81ea\u4e3b\u5904\u7406\u4efb\u52a1\u3001\u8bbe\u8ba1\u5b9e\u9a8c\u3001\u57fa\u4e8e\u5386\u53f2\u8fed\u4ee3\u4f18\u5316<\/li><li>12 \u4e2a ML \u4efb\u52a1\u4e0a\u5339\u914d\/\u8d85\u8d8a\u4eba\u7c7b\u6700\u4f73\u8bd5\u9a8c<\/li><li>\u63d0\u4f9b\u53ef\u89e3\u91ca\u7ed3\u679c<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4efb\u4f55\u9700\u8981\u8d85\u53c2\u6570\u8c03\u4f18\u7684 ML \u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"52-automl-agent--\">5.2 AutoML-Agent \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/DeepAuto-AI\/automl-agent\">https:\/\/github.com\/DeepAuto-AI\/automl-agent<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: ICML 2025<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u591a Agent \u534f\u4f5c\u7684\u5168\u6d41\u7a0b AutoML<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u5e76\u884c\u5316\u4e13\u7528 Agent \u5206\u5de5\u534f\u4f5c<\/li><li>\u68c0\u7d22\u589e\u5f3a\u89c4\u5212<\/li><li>\u591a\u9636\u6bb5\u9a8c\u8bc1<\/li><li>14 \u4e2a\u6570\u636e\u96c6\u9a8c\u8bc1<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5168\u81ea\u52a8\u5316 ML Pipeline<\/li><\/ul>\n\n\n\n<h3 id=\"53-optuna-\">5.3 Optuna \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/optuna.org\">https:\/\/optuna.org<\/a><\/li><li><strong>\u5b9a\u4f4d<\/strong>: \u8d85\u53c2\u6570\u4f18\u5316\u6846\u67b6\uff08\u884c\u4e1a\u6807\u51c6\uff09<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>: Bayesian \u641c\u7d22\u3001\u526a\u679d\u3001\u53ef\u89c6\u5316 Dashboard\u3001\u5206\u5e03\u5f0f\u6267\u884c<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u8bad\u7ec3\u6d41\u7a0b\u4e2d\u7684\u6807\u51c6\u8d85\u53c2\u6570\u641c\u7d22\u7ec4\u4ef6<\/li><\/ul>\n\n\n\n<h3 id=\"54-microsoft-nni-\">5.4 Microsoft NNI \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/microsoft\/nni\">https:\/\/github.com\/microsoft\/nni<\/a><\/li><li><strong>\u6838\u5fc3\u529f\u80fd<\/strong>: AutoML \u5168\u751f\u547d\u5468\u671f<ul><li><strong>\u7279\u5f81\u5de5\u7a0b\u81ea\u52a8\u5316<\/strong><\/li><li><strong>\u795e\u7ecf\u67b6\u6784\u641c\u7d22\uff08NAS\uff09<\/strong><\/li><li><strong>\u8d85\u53c2\u6570\u8c03\u4f18<\/strong><\/li><li><strong>\u6a21\u578b\u538b\u7f29<\/strong><\/li><\/ul><\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u652f\u6301\u672c\u5730\/\u8fdc\u7a0b\/\u4e91\u591a\u79cd\u73af\u5883<\/li><li>Web UI \u7ba1\u7406\u5b9e\u9a8c<\/li><li>\u53ef\u6269\u5c55 API \u81ea\u5b9a\u4e49\u7b97\u6cd5<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4f01\u4e1a\u7ea7 AutoML\u3001NAS\u3001\u6a21\u578b\u538b\u7f29<\/li><\/ul>\n\n\n\n<h3 id=\"55-wb-sweeps-\">5.5 W&amp;B Sweeps \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/wandb.ai\/site\/sweeps\">https:\/\/wandb.ai\/site\/sweeps<\/a><\/li><li><strong>\u6838\u5fc3\u529f\u80fd<\/strong>: \u81ea\u52a8\u5316\u8d85\u53c2\u6570\u641c\u7d22 + \u5b9e\u9a8c\u8ffd\u8e2a<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>Bayesian \/ Grid \/ Random \u641c\u7d22<\/li><li>Hyperband \u65e9\u505c\uff1a\u81ea\u52a8\u6740\u6389\u5dee\u7684\u5b9e\u9a8c\uff0c\u91ca\u653e GPU<\/li><li>\u53c2\u6570\u91cd\u8981\u6027\u5206\u6790<\/li><li>\u53ef\u8de8\u591a\u673a\u5e76\u884c<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4e0e W&amp;B \u5b9e\u9a8c\u8ffd\u8e2a\u6df1\u5ea6\u96c6\u6210\u7684\u8d85\u53c2\u6570\u4f18\u5316<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"6-\u81ea\u8fdb\u5316--self-play-\u8bad\u7ec3\u65b9\u6cd5\">6. \u81ea\u8fdb\u5316 \/ Self-Play \u8bad\u7ec3\u65b9\u6cd5<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u6838\u5fc3\u601d\u8def\uff1a<strong>\u6a21\u578b\u81ea\u5df1\u751f\u6210\u8bad\u7ec3\u6570\u636e\u6765\u8bad\u7ec3\u81ea\u5df1<\/strong>\uff0c\u51cf\u5c11\u5bf9\u4eba\u5de5\u6807\u6ce8\u7684\u4f9d\u8d56\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"61-spin-self-play-fine-tuning-\">6.1 SPIN (Self-Play Fine-Tuning) \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/uclaml\/SPIN\">https:\/\/github.com\/uclaml\/SPIN<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2401.01335\">https:\/\/arxiv.org\/abs\/2401.01335<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u6a21\u578b\u4e0e\u81ea\u5df1\u7684\u5386\u53f2\u7248\u672c\u5bf9\u5f08\uff0c\u533a\u5206\u81ea\u751f\u6210\u56de\u590d\u548c SFT \u6570\u636e<\/li><li><strong>\u5173\u952e\u7ed3\u679c<\/strong>: \u65e0\u9700\u989d\u5916\u4eba\u5de5\u504f\u597d\u6570\u636e\uff0c\u6027\u80fd\u8d85\u8d8a DPO + GPT-4 \u504f\u597d\u6570\u636e<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4f4e\u6210\u672c\u5bf9\u9f50\u8bad\u7ec3\u3001\u81ea\u6211\u6539\u8fdb\u5faa\u73af<\/li><\/ul>\n\n\n\n<h3 id=\"62-sppo-self-play-preference-optimization-\">6.2 SPPO (Self-Play Preference Optimization) \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u6765\u6e90<\/strong>: UCLA ML Lab<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u901a\u8fc7\u8fed\u4ee3\u7b56\u7565\u66f4\u65b0\u903c\u8fd1 Nash \u5747\u8861<\/li><li><strong>\u7279\u70b9<\/strong>: \u6709\u7406\u8bba\u6536\u655b\u4fdd\u8bc1<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u504f\u597d\u5bf9\u9f50\u7684\u7406\u8bba\u9a71\u52a8\u65b9\u6cd5<\/li><\/ul>\n\n\n\n<h3 id=\"63-spc-self-play-critic--\">6.3 SPC (Self-Play Critic) \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/chen-judge.github.io\/SPC\/\">https:\/\/chen-judge.github.io\/SPC\/<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u5bf9\u6297\u5f0f\u81ea\u5bf9\u5f08\u8fdb\u5316\u63a8\u7406\u8bc4\u5224\u8005<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>&#8220;\u72e1\u733e\u751f\u6210\u5668&#8221; vs &#8220;\u8bc4\u5224\u8005&#8221; \u535a\u5f08\u673a\u5236<\/li><li>\u6d88\u9664\u624b\u52a8\u6b65\u9aa4\u7ea7\u6807\u6ce8\u9700\u6c42<\/li><li>\u63d0\u5347\u63a8\u7406\u9a8c\u8bc1\u80fd\u529b<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u8bad\u7ec3\u66f4\u5f3a\u7684\u63a8\u7406\u9a8c\u8bc1\/\u8bc4\u5224\u6a21\u578b<\/li><\/ul>\n\n\n\n<h3 id=\"64-spell--\">6.4 SPELL \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/html\/2509.23863\">https:\/\/arxiv.org\/html\/2509.23863<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u81ea\u5bf9\u5f08 RL \u8fdb\u5316\u957f\u4e0a\u4e0b\u6587\u8bed\u8a00\u6a21\u578b<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u65e0\u6807\u7b7e\u81ea\u5bf9\u5f08\u65b9\u6cd5<\/li><li>\u57fa\u7840\u6a21\u578b\u5728\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u4e0a\u8d85\u8d8a\u7ecf\u8fc7\u6307\u4ee4\u5fae\u8c03\u7684\u7248\u672c<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u957f\u4e0a\u4e0b\u6587\u80fd\u529b\u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"65-multi-agent-evolve-mae-\">6.5 Multi-Agent Evolve (MAE) \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/html\/2510.23595v1\">https:\/\/arxiv.org\/html\/2510.23595v1<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u4e00\u4e2a LLM \u540c\u65f6\u626e\u6f14 Proposer\u3001Solver\u3001Judge \u591a\u4e2a\u89d2\u8272\uff0c\u534f\u540c\u8fdb\u5316<\/li><li><strong>\u9a8c\u8bc1<\/strong>: Qwen2.5-3B-Instruct \u5728\u6570\u5b66\u3001\u7f16\u7a0b\u3001\u63a8\u7406\u3001\u901a\u8bc6\u4e0a\u5747\u6709\u63d0\u5347<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u591a\u89d2\u8272\u81ea\u8fdb\u5316\u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"66-multiagent-finetuning-\">6.6 Multiagent Finetuning \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/llm-multiagent-ft.github.io\/\">https:\/\/llm-multiagent-ft.github.io\/<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u4ece\u540c\u4e00 base model \u521d\u59cb\u5316\u591a\u4e2a\u6a21\u578b\uff0c\u901a\u8fc7\u591a Agent \u4ea4\u4e92\u751f\u6210\u6570\u636e\uff0c\u72ec\u7acb\u66f4\u65b0<\/li><li><strong>\u5173\u952e\u53d1\u73b0<\/strong>: \u591a Agent \u5fae\u8c03\u53ef\u6301\u7eed\u591a\u8f6e\u8fed\u4ee3\u6539\u8fdb\uff0c\u800c\u5355\u6a21\u578b\u81ea\u8bad\u7ec3\u4f1a plateau<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u514b\u670d\u81ea\u8bad\u7ec3\u74f6\u9888<\/li><\/ul>\n\n\n\n<h3 id=\"67-cory-cooperative-multi-agent-rl-fine-tuning-\">6.7 CORY (Cooperative Multi-Agent RL Fine-Tuning) \u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u6765\u6e90<\/strong>: NeurIPS 2024<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u5c06 LLM \u590d\u5236\u4e3a Pioneer + Observer \u4e24\u4e2a Agent\uff0c\u534f\u4f5c RL \u5fae\u8c03<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: RL \u5fae\u8c03\u7684\u65b0\u8303\u5f0f<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"7-\u5408\u6210\u6570\u636e\u751f\u6210\u4e0e\u6570\u636e\u6cbb\u7406-\">7. \u5408\u6210\u6570\u636e\u751f\u6210\u4e0e\u6570\u636e\u6cbb\u7406 \ud83c\udd95<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u81ea\u52a8\u5316\u8bad\u7ec3 Pipeline \u7684\u5173\u952e\u73af\u8282\uff1a<strong>\u5927\u89c4\u6a21\u751f\u6210\u9ad8\u8d28\u91cf\u8bad\u7ec3\u6570\u636e<\/strong>\uff0c\u65e0\u9700\u4eba\u5de5\u6807\u6ce8\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"71-\u6570\u636e\u751f\u6210\u6846\u67b6\">7.1 \u6570\u636e\u751f\u6210\u6846\u67b6<\/h3>\n\n\n\n<h4 id=\"distilabel-\">Distilabel \u2b50\u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/argilla-io\/distilabel\">https:\/\/github.com\/argilla-io\/distilabel<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: Argilla<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u57fa\u4e8e\u5df2\u9a8c\u8bc1\u7814\u7a76\u8bba\u6587\u7684\u5408\u6210\u6570\u636e\u548c AI \u53cd\u9988 Pipeline<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u6a21\u5757\u5316 Pipeline \u67b6\u6784\uff08\u53ef\u94fe\u5f0f\u7ec4\u5408\u7684 Steps\/Tasks\uff09<\/li><li>\u652f\u6301 SFT\u3001DPO\u3001UltraFeedback\u3001DEITA \u7b49\u6280\u672f<\/li><li>\u7edf\u4e00 API \u5bf9\u63a5\u4efb\u4f55 LLM \u63d0\u4f9b\u5546<\/li><li>\u5185\u7f6e Argilla \u96c6\u6210\u7528\u4e8e\u6807\u6ce8\u548c\u5ba1\u6838<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u89c4\u6a21\u5316\u751f\u6210 SFT\/DPO\/\u504f\u597d\u6570\u636e<\/li><\/ul>\n\n\n\n<h4 id=\"magpie-\">Magpie \u2b50\u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/magpie-align\/magpie\">https:\/\/github.com\/magpie-align\/magpie<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: ICLR 2025<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u4ece\u96f6\u5f00\u59cb\u5408\u6210\u5bf9\u9f50\u6570\u636e\uff0c\u65e0\u9700 Prompt Engineering \u6216\u79cd\u5b50\u95ee\u9898<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u5229\u7528\u5bf9\u9f50 LLM \u7684\u9884\u67e5\u8be2\u6a21\u677f\u81ea\u52a8\u751f\u6210<\/li><li>\u751f\u6210 4M \u6761\u6307\u4ee4\uff0c\u7b5b\u9009\u540e 300K \u9ad8\u8d28\u91cf\u6837\u672c<\/li><li>\u5728 Magpie \u6570\u636e\u4e0a\u5fae\u8c03\u7684\u6a21\u578b\u5339\u914d\u5b98\u65b9 Llama-3-8B-Instruct<\/li><li>Magpie Reasoning V2: 250K \u6761 CoT \u63a8\u7406\u6837\u672c<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4f4e\u6210\u672c\u5927\u89c4\u6a21\u5bf9\u9f50\u6570\u636e\u751f\u6210<\/li><\/ul>\n\n\n\n<h4 id=\"datadreamer-\">DataDreamer \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/datadreamer-dev\/DataDreamer\">https:\/\/github.com\/datadreamer-dev\/DataDreamer<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: ACL 2024<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u53ef\u590d\u73b0\u7684\u5408\u6210\u6570\u636e\u751f\u6210\u4e0e LLM \u5de5\u4f5c\u6d41<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u591a\u6b65 Prompting \u5de5\u4f5c\u6d41<\/li><li>\u751f\u6210\u6570\u636e + \u5bf9\u9f50 + \u5fae\u8c03 + \u84b8\u998f\u4e00\u4f53\u5316<\/li><li>\u5185\u7f6e\u53ef\u590d\u73b0\u6027\u548c\u7f13\u5b58\u673a\u5236<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u53ef\u590d\u73b0\u7684\u6570\u636e\u751f\u6210\u7814\u7a76<\/li><\/ul>\n\n\n\n<h4 id=\"cosmopedia-\">Cosmopedia \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/huggingface\/cosmopedia\">https:\/\/github.com\/huggingface\/cosmopedia<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: HuggingFace<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u5927\u89c4\u6a21\u5408\u6210\u9884\u8bad\u7ec3\u6570\u636e Pipeline<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u4f7f\u7528 Mixtral-8x7B \u751f\u6210 <strong>25B tokens<\/strong> \u7684\u5408\u6210\u6559\u79d1\u4e66\u3001\u535a\u5ba2\u3001\u6545\u4e8b<\/li><li>\u4f7f\u7528 Stanford\u3001Khan Academy\u3001OpenStax\u3001WikiHow \u7b49\u8bdd\u9898<\/li><li>\u6700\u5927\u7684\u5f00\u653e\u5408\u6210\u6570\u636e\u96c6<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9884\u8bad\u7ec3\u6570\u636e\u589e\u5f3a<\/li><\/ul>\n\n\n\n<h4 id=\"instructlab-sdg-\">InstructLab SDG \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/instructlab\/sdg\">https:\/\/github.com\/instructlab\/sdg<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: IBM \/ Red Hat<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u57fa\u4e8e LAB \u65b9\u6cd5\u8bba\u7684\u5408\u6210\u6570\u636e\u751f\u6210<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>Skills-SDG + Knowledge-SDG \u53cc\u751f\u6210\u5668<\/li><li>\u4ece\u5c11\u91cf\u79cd\u5b50\u5206\u7c7b\u5b66\u81ea\u52a8\u6269\u5c55\u5230\u5927\u89c4\u6a21\u8bad\u7ec3\u6570\u636e<\/li><li>\u793e\u533a\u8d21\u732e\u673a\u5236<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u793e\u533a\u9a71\u52a8\u7684\u6a21\u578b\u80fd\u529b\u6269\u5c55<\/li><\/ul>\n\n\n\n<h4 id=\"persona-hub-\">Persona Hub \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/tencent-ailab\/persona-hub\">https:\/\/github.com\/tencent-ailab\/persona-hub<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u817e\u8baf AI Lab<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u57fa\u4e8e\u89d2\u8272\u9a71\u52a8\u7684\u5341\u4ebf\u7ea7\u5408\u6210\u6570\u636e<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u4ece\u7f51\u7edc\u6570\u636e\u4e2d\u63d0\u53d6 <strong>10 \u4ebf\u4e2a\u591a\u6837\u5316\u89d2\u8272<\/strong><\/li><li>2025-02 \u53d1\u5e03 3.7 \u4ebf &#8220;\u7cbe\u82f1&#8221; \u89d2\u8272<\/li><li>\u9a71\u52a8\u591a\u6837\u5316\u5408\u6210\u6570\u636e\u751f\u6210<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u9ad8\u591a\u6837\u6027\u8bad\u7ec3\u6570\u636e\u7684\u573a\u666f<\/li><\/ul>\n\n\n\n<h4 id=\"synthgen-\">synth_gen \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/facebookresearch\/synth_gen\">https:\/\/github.com\/facebookresearch\/synth_gen<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: Meta<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u5e26\u6267\u884c\u9a8c\u8bc1\u7684\u5408\u6210\u6570\u636e\u751f\u6210<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u6a21\u5757\u5316\u9a8c\u8bc1\u5668\u7cfb\u7edf\uff08\u68c0\u67e5\u5668\u3001\u9519\u8bef\u6d88\u606f\u3001\u4fee\u590d\u5668\uff09<\/li><li>\u4ece\u79cd\u5b50\u751f\u6210\u5df2\u9a8c\u8bc1\u7684\u8bad\u7ec3\u5bf9\u8bdd<\/li><li>\u57fa\u4e8e\u89e3\u6790\u5668\u7684\u4ee3\u7801\u9a8c\u8bc1<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u9ad8\u51c6\u786e\u6027\u7684\u4ee3\u7801\/\u63a8\u7406\u6570\u636e\u751f\u6210<\/li><\/ul>\n\n\n\n<h4 id=\"nemo-data-designer--\">NeMo Data Designer \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/NVIDIA-NeMo\/DataDesigner\">https:\/\/github.com\/NVIDIA-NeMo\/DataDesigner<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: NVIDIA NeMo \u5b98\u65b9\uff08v0.5.1\uff0c2026-02\uff09<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u8d85\u8d8a\u7b80\u5355 LLM \u63d0\u793a\u7684\u751f\u4ea7\u7ea7\u5408\u6210\u6570\u636e\u751f\u6210\u7cfb\u7edf\u2014\u2014\u7edf\u8ba1\u91c7\u6837\u5668 + LLM \u751f\u6210 + \u4f9d\u8d56\u611f\u77e5\u5b57\u6bb5 + \u81ea\u52a8\u9a8c\u8bc1<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u4f9d\u8d56\u611f\u77e5\u751f\u6210<\/strong>\uff1a\u5b57\u6bb5\u95f4\u53ef\u5b9a\u4e49\u5173\u8054\u5173\u7cfb\uff0c\u4fdd\u6301\u6570\u636e\u5185\u5728\u4e00\u81f4\u6027<\/li><li>\u652f\u6301\u7edf\u8ba1\u91c7\u6837\u5668\u3001LLM \u751f\u6210\u5668\u548c\u5df2\u6709\u79cd\u5b50\u6570\u636e\u96c6\u4e09\u79cd\u6765\u6e90<\/li><li>\u5185\u7f6e\u9a8c\u8bc1\u5668\uff08Python\u3001SQL\u3001\u81ea\u5b9a\u4e49\uff09+ <strong>LLM-as-judge \u8d28\u91cf\u8bc4\u5206<\/strong><\/li><li>\u652f\u6301<strong>\u56fe\u50cf\u751f\u6210<\/strong>\uff08v0.5.1\uff09\uff0c\u53ef\u6784\u5efa\u591a\u6a21\u6001\u5408\u6210\u6570\u636e\u96c6<\/li><li>Preview \u6a21\u5f0f\u5feb\u901f\u8fed\u4ee3\uff0c\u6ee1\u610f\u540e\u518d\u5168\u91cf\u751f\u6210<\/li><li>\u652f\u6301 Docker Compose \u6216 Helm \u90e8\u7f72<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u7ed3\u6784\u5316\u3001\u9ad8\u8d28\u91cf\u3001\u53ef\u9a8c\u8bc1\u5408\u6210\u6570\u636e\u7684\u5927\u89c4\u6a21\u8bad\u7ec3 Pipeline<\/li><\/ul>\n\n\n\n<h4 id=\"evidently--\">Evidently \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/evidentlyai\/evidently\">https:\/\/github.com\/evidentlyai\/evidently<\/a><\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u5f00\u6e90\u5408\u6210\u6570\u636e\u751f\u6210\uff0c\u652f\u6301\u7528\u6237\u753b\u50cf\u4e0e\u65e0\u4ee3\u7801\u754c\u9762<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u53ef\u5b9a\u4e49\u7528\u6237\u753b\u50cf\uff08\u8bed\u6c14\u3001\u610f\u56fe\u3001\u77e5\u8bc6\u6c34\u5e73\u3001&#8221;\u602a\u7656&#8221;\uff09\u6765\u5851\u9020\u751f\u6210\u8f93\u5165<\/li><li>\u6a21\u578b\u65e0\u5173\uff1a\u53ef\u5bf9\u63a5\u4efb\u4f55 LLM \u63d0\u4f9b\u5546<\/li><li>\u8f93\u51fa\u76f4\u63a5\u5230 pandas DataFrame<\/li><li>Evidently Cloud \u63d0\u4f9b\u65e0\u4ee3\u7801 UI\uff0c\u4ea7\u54c1\u7ecf\u7406\u548c\u9886\u57df\u4e13\u5bb6\u53ef\u76f4\u63a5\u4f7f\u7528<\/li><li>\u5185\u7f6e LLM \u8bc4\u4f30\u6307\u6807\u3001\u5408\u6210\u6570\u636e\u751f\u6210\u3001\u81ea\u52a8 Prompt \u4f18\u5316<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: LLM \u7cfb\u7edf\u6d4b\u8bd5\u6570\u636e\u751f\u6210\u3001\u8bc4\u4f30\u6570\u636e\u96c6\u6784\u5efa<\/li><\/ul>\n\n\n\n<h4 id=\"meta-synthetic-data-kit--\">Meta Synthetic-Data-Kit \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/meta-llama\/synthetic-data-kit\">https:\/\/github.com\/meta-llama\/synthetic-data-kit<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: Meta\uff08meta-llama \u5b98\u65b9\u7ec4\u7ec7\uff09<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u8f7b\u91cf\u7ea7 CLI \u5de5\u5177\uff0c\u4ece\u4efb\u610f\u6587\u6863\u5feb\u901f\u751f\u6210\u9ad8\u8d28\u91cf LLM \u5fae\u8c03\u6570\u636e\u96c6\uff0cLlama 3 \u7684\u5927\u90e8\u5206 SFT \u8bed\u6599\u5373\u901a\u8fc7 6 \u8f6e\u8fed\u4ee3\u5408\u6210\u751f\u6210<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u56db\u6b65\u6d41\u6c34\u7ebf CLI<\/strong>\uff1a<code>ingest<\/code>\uff08\u89e3\u6790\u6587\u6863\uff09\u2192 <code>create<\/code>\uff08\u751f\u6210\u6570\u636e\uff09\u2192 <code>curate<\/code>\uff08\u8d28\u91cf\u8fc7\u6ee4\uff09\u2192 <code>save-as<\/code>\uff08\u683c\u5f0f\u8f6c\u6362\uff09<\/li><li>\u652f\u6301\u591a\u79cd\u8f93\u5165\u683c\u5f0f\uff1aPDF\u3001HTML\u3001YouTube \u5b57\u5e55\u3001DOCX\u3001PPT\u3001TXT<\/li><li>\u751f\u6210\u6570\u636e\u7c7b\u578b\uff1aQA \u5bf9\u3001CoT \u63a8\u7406\u94fe\u3001\u6458\u8981\uff1b\u652f\u6301\u81ea\u5b9a\u4e49 Prompt \u6a21\u677f<\/li><li><strong>LLM-as-judge<\/strong>\uff1a\u540c\u4e00 LLM \u65e2\u8d1f\u8d23\u751f\u6210\u4e5f\u8d1f\u8d23\u8d28\u91cf\u8fc7\u6ee4\uff0c\u53ef\u914d\u7f6e\u8d28\u91cf\u9608\u503c<\/li><li>\u5bfc\u51fa\u683c\u5f0f\uff1aJSONL\u3001Alpaca\u3001ChatML\u3001OpenAI\uff0c\u8986\u76d6\u4e3b\u6d41\u8bad\u7ec3\u6846\u67b6<\/li><li>\u57fa\u4e8e vLLM \u8fd0\u884c\uff0c\u53ef\u5bf9\u63a5 Llama\u3001Qwen \u7b49\u672c\u5730\u6a21\u578b<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4ece\u5185\u90e8\u6587\u6863\/\u8bba\u6587\/\u4ee3\u7801\u5e93\u5feb\u901f\u6784\u5efa\u9886\u57df\u4e13\u5c5e\u5fae\u8c03\u6570\u636e\u96c6\uff1b\u65e0\u9700\u5916\u90e8\u6570\u636e\u6807\u6ce8\u56e2\u961f<\/li><\/ul>\n\n\n\n<h4 id=\"dataflow--\">DataFlow \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/OpenDCAI\/DataFlow\">https:\/\/github.com\/OpenDCAI\/DataFlow<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2512.16676\uff08DataFlow:\">https:\/\/arxiv.org\/abs\/2512.16676\uff08DataFlow:<\/a> An LLM-Driven Framework for Unified Data Preparation and Workflow Automation\uff09<\/li><li><strong>\u6765\u6e90<\/strong>: OpenDCAI<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: LLM \u9a71\u52a8\u7684\u7edf\u4e00\u6570\u636e\u51c6\u5907\u4e0e\u5de5\u4f5c\u6d41\u81ea\u52a8\u5316\u6846\u67b6\u2014\u2014\u7528\u300c\u53ef\u7ec4\u5408\u62bd\u8c61 + LLM \u4f18\u5148\u7684\u7b97\u5b50\u6267\u884c\u6a21\u578b\u300d\u628a\u6570\u636e\u51c6\u5907\u672c\u8eab\u53d8\u6210\u53ef\u81ea\u52a8\u7f16\u6392\u7684\u6d41\u6c34\u7ebf<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u8fd1 200 \u4e2a\u53ef\u590d\u7528\u7b97\u5b50<\/strong> + 6 \u6761 SOTA \u6a21\u677f\u6d41\u6c34\u7ebf\uff1a\u6587\u672c\u3001\u6570\u5b66\u63a8\u7406\u3001\u4ee3\u7801\u3001Text-to-SQL\u3001Agentic RAG \u6570\u636e\u3001\u5927\u89c4\u6a21 QA \u62bd\u53d6<\/li><li><strong>DataFlow-Agent<\/strong>\uff1a\u628a\u81ea\u7136\u8bed\u8a00\u9700\u6c42\u81ea\u52a8\u7ffb\u8bd1\u6210\u53ef\u6267\u884c\u6d41\u6c34\u7ebf\u2014\u2014\u7b97\u5b50\u5408\u6210 \u2192 \u6d41\u6c34\u7ebf\u89c4\u5212 \u2192 \u8fed\u4ee3\u9a8c\u8bc1\uff1b\u4e5f\u53ef\u6309\u4efb\u52a1\u76ee\u6807\u81ea\u884c\u7f16\u5199\u81ea\u5b9a\u4e49\u7b97\u5b50<\/li><li><strong>DataFlow-WebUI<\/strong>\uff1a\u62d6\u62fd\u5f0f\u6d41\u6c34\u7ebf\u7f16\u8f91 + \u81ea\u7136\u8bed\u8a00\u4ea4\u4e92\u7684\u53ef\u89c6\u5316\u5de5\u4f5c\u53f0<\/li><li>\u914d\u5957 DataFlow-Skills\uff0c\u53ef\u88ab\u7f16\u7801 Agent \u76f4\u63a5\u8c03\u7528<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u628a\u300c\u6570\u636e\u6e05\u6d17 + \u5408\u6210 + \u8bc4\u4f30\u300d\u7edf\u4e00\u7f16\u6392\u7684\u8bad\u7ec3\u524d\u7f6e\u6d41\u7a0b\uff1b\u7528\u81ea\u7136\u8bed\u8a00\u9a71\u52a8\u6570\u636e\u51c6\u5907\u6d41\u6c34\u7ebf<\/li><\/ul>\n\n\n\n<h3 id=\"72-\u6570\u636e\u6cbb\u7406\u4e0e\u7b5b\u9009\">7.2 \u6570\u636e\u6cbb\u7406\u4e0e\u7b5b\u9009<\/h3>\n\n\n\n<h4 id=\"nemo-curator-\">NeMo Curator \u2b50\u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/NVIDIA-NeMo\/Curator\">https:\/\/github.com\/NVIDIA-NeMo\/Curator<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: NVIDIA<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: GPU \u52a0\u901f\u7684\u5927\u89c4\u6a21\u6570\u636e\u9884\u5904\u7406\u4e0e\u6cbb\u7406<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>30+ \u542f\u53d1\u5f0f\u8fc7\u6ee4\u5668<\/li><li>fastText + GPU \u52a0\u901f\u5206\u7c7b\u5668\uff08\u9886\u57df\/\u8d28\u91cf\/\u5b89\u5168\uff09<\/li><li>\u6a21\u7cca\u53bb\u91cd\uff1a64 \u5f20 A100 \u4e0a 1.8 \u5c0f\u65f6\u5904\u7406 1.1T tokens<\/li><li>\u652f\u6301\u6587\u672c\u3001\u56fe\u50cf\u3001\u89c6\u9891\u3001\u97f3\u9891<\/li><li>\u6bd4\u66ff\u4ee3\u65b9\u6848\u5feb 16x<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u6570\u636e\u6e05\u6d17<\/li><\/ul>\n\n\n\n<h4 id=\"datatrove-\">DataTrove \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/huggingface\/datatrove\">https:\/\/github.com\/huggingface\/datatrove<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: HuggingFace<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u5e73\u53f0\u65e0\u5173\u7684\u6570\u636e\u5904\u7406 Pipeline<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u6a21\u5757\u5316\u8bfb\u53d6\u5668\/\u5199\u5165\u5668\/\u63d0\u53d6\u5668\/\u8fc7\u6ee4\u5668<\/li><li>\u672c\u5730\u6216 Slurm \u96c6\u7fa4\u8fd0\u884c<\/li><li>\u4f4e\u5185\u5b58\u5360\u7528<\/li><li>\u7528\u4e8e FineWeb \u548c Cosmopedia \u6570\u636e\u96c6<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5927\u89c4\u6a21\u6587\u672c\u6570\u636e\u5904\u7406<\/li><\/ul>\n\n\n\n<h4 id=\"dolma-\">Dolma \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/allenai\/dolma\">https:\/\/github.com\/allenai\/dolma<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: AllenAI<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u9ad8\u6027\u80fd\u6570\u636e\u96c6\u6cbb\u7406\u5de5\u5177<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u652f\u6301\u6570\u5341\u4ebf\u6587\u6863\u7684\u5185\u7f6e\u5e76\u884c\u5904\u7406<\/li><li>CommonCrawl \u4e0b\u8f7d\u3001\u8fc7\u6ee4\u3001\u53bb\u91cd\u3001\u5206\u6790<\/li><li>\u8bed\u8a00\u68c0\u6d4b\u3001PII \u6e05\u7406<\/li><li>\u7528\u4e8e OLMo \u8bad\u7ec3\u8bed\u6599\uff083T tokens\uff09<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9884\u8bad\u7ec3\u8bed\u6599\u6cbb\u7406<\/li><\/ul>\n\n\n\n<h4 id=\"data-prep-kit-\">Data Prep Kit \u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/data-prep-kit\/data-prep-kit\">https:\/\/github.com\/data-prep-kit\/data-prep-kit<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: IBM<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u975e\u7ed3\u6784\u5316\u6570\u636e\u51c6\u5907<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>Python \/ Ray \/ Spark \u8fd0\u884c\u65f6<\/li><li>\u4ece\u7b14\u8bb0\u672c\u5230\u6570\u636e\u4e2d\u5fc3\u6269\u5c55<\/li><li>Kubeflow Pipelines \u5de5\u4f5c\u6d41\u81ea\u52a8\u5316<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u591a\u6a21\u6001\u6570\u636e\u51c6\u5907<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"8-\u77e5\u8bc6\u84b8\u998f\u6846\u67b6-\">8. \u77e5\u8bc6\u84b8\u998f\u6846\u67b6 \ud83c\udd95<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p><strong>\u5c06\u5927\u6a21\u578b\u7684\u80fd\u529b\u538b\u7f29\u5230\u5c0f\u6a21\u578b\u4e2d<\/strong>\uff0c\u4fdd\u6301\u6027\u80fd\u7684\u540c\u65f6\u964d\u4f4e\u90e8\u7f72\u6210\u672c\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"81-easydistill-\">8.1 EasyDistill \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/modelscope\/easydistill\">https:\/\/github.com\/modelscope\/easydistill<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u963f\u91cc\u5df4\u5df4 \/ ModelScope\uff08EMNLP 2025 Demo\uff09<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u7efc\u5408\u84b8\u998f\u5de5\u5177\u5305\uff0c\u9ed1\u76d2 + \u767d\u76d2\u65b9\u6cd5\u5168\u8986\u76d6<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u6570\u636e\u5408\u6210 + SFT + Logits \u84b8\u998f + \u6392\u5e8f\u4f18\u5316 + RL for KD<\/li><li>\u5305\u542b\u84b8\u998f\u6a21\u578b\uff08DistilQwen\uff09\u548c\u5f00\u6e90\u6570\u636e\u96c6<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u751f\u4ea7\u7ea7\u6a21\u578b\u84b8\u998f<\/li><\/ul>\n\n\n\n<h3 id=\"82-distillkit-\">8.2 DistillKit \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/arcee-ai\/DistillKit\">https:\/\/github.com\/arcee-ai\/DistillKit<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: Arcee AI<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u751f\u4ea7\u7ea7 LLM \u84b8\u998f\u5de5\u5177<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u5728\u7ebf\u548c\u79bb\u7ebf\u84b8\u998f\u5de5\u4f5c\u6d41<\/li><li>\u9a71\u52a8 Arcee Virtuoso\u3001SuperNova Medius \u7b49\u6a21\u578b<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u751f\u4ea7\u90e8\u7f72\u524d\u7684\u6a21\u578b\u538b\u7f29<\/li><\/ul>\n\n\n\n<h3 id=\"83-miniplm-\">8.3 MiniPLM \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/thu-coai\/MiniPLM\">https:\/\/github.com\/thu-coai\/MiniPLM<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u6e05\u534e\u5927\u5b66\uff08ICLR 2025\uff09<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u9762\u5411\u9884\u8bad\u7ec3\u7684\u77e5\u8bc6\u84b8\u998f<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9ad8\u6548\u9884\u8bad\u7ec3\u5c0f\u6a21\u578b<\/li><\/ul>\n\n\n\n<h3 id=\"84-distillm-\">8.4 DistiLLM \u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/jongwooko\/distillm\">https:\/\/github.com\/jongwooko\/distillm<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: ICML 2024<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u5bf9\u6bd4\u5f0f\u84b8\u998f\u65b9\u6cd5<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u7814\u7a76\u7ea7\u84b8\u998f\u5b9e\u9a8c<\/li><\/ul>\n\n\n\n<h3 id=\"85-easyopd--\">8.5 EasyOPD \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2607.11012\uff08EasyOPD:\">https:\/\/arxiv.org\/abs\/2607.11012\uff08EasyOPD:<\/a> An Easy-to-use On-Policy Distillation Framework for Large Language Models\uff09<\/li><li><strong>\u53d1\u5e03<\/strong>: 2026-07-13<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u9996\u4e2a\u7edf\u4e00\u7684<strong>\u5728\u7ebf\u7b56\u7565\u84b8\u998f\uff08On-Policy Distillation, OPD\uff09<\/strong>\u6846\u67b6\u2014\u2014\u5b66\u751f\u6a21\u578b\u81ea\u5df1\u751f\u6210\u3001\u6559\u5e08\u6a21\u578b\u5728\u7ebf\u6253\u5206\uff0c\u628a OPD \u505a\u5f97\u50cf\u5fae\u8c03\u4e00\u6837\u53ef\u914d\u7f6e<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u8986\u76d6 <strong>10+ \u79cd OPD \u65b9\u6cd5<\/strong>\uff0c\u901a\u8fc7<strong>\u4e00\u884c YAML \u5207\u6362<\/strong>\uff1a\u8de8 tokenizer OPD\u3001\u5728\u7ebf\u81ea\u84b8\u998f\uff08self-distillation\uff09\u3001\u6b65\u9aa4\u7ea7\uff08step-wise\uff09OPD \u4e09\u5927\u8bbe\u5b9a<\/li><li>\u57fa\u4e8e <strong>verl<\/strong> \u5206\u5e03\u5f0f RL \u540e\u7aef\u6784\u5efa\uff0c\u5c06\u7528\u6237\u4fa7\u914d\u7f6e\u3001\u65b9\u6cd5\u7279\u5b9a\u7684\u76d1\u7763\u903b\u8f91\u4e0e\u6267\u884c\u5c42\u89e3\u8026<\/li><li>\u5404\u65b9\u6cd5\u6a21\u5757\u901a\u8fc7\u7edf\u4e00\u6269\u5c55\u8fb9\u754c\u63a5\u5165\u5171\u4eab\u540e\u7aef\uff1a\u635f\u5931\u6784\u9020\u3001rollout \u5143\u6570\u636e\u3001\u5956\u52b1\u5904\u7406\u3001tokenizer \u5bf9\u9f50\u3001\u6559\u5e08\u4fa7\u8ba1\u7b97<\/li><li>\u5728\u63a8\u7406\u3001\u4ee3\u7801\u751f\u6210\u3001\u79d1\u5b66\u77e5\u8bc6\u3001\u5de5\u5177\u8c03\u7528\u7b49\u591a\u7c7b\u57fa\u51c6\u4e0a\u9a8c\u8bc1<\/li><li>\u968f\u9644\u53ef\u8fd0\u884c YAML \u914d\u7f6e\u3001\u6587\u6863\u3001\u53ef\u5b89\u88c5\u7684\u6f14\u793a\u5305\u4e0e\u89c6\u9891<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u628a\u5728\u7ebf\u7b56\u7565\u84b8\u998f\u4f5c\u4e3a SFT \u4e0e RL \u4e4b\u95f4\u7684\u6807\u51c6\u540e\u8bad\u7ec3\u9636\u6bb5\uff1b\u7cfb\u7edf\u5bf9\u6bd4\u591a\u79cd OPD \u65b9\u6cd5<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"9-\u6a21\u578b\u5408\u5e76\u4e0e\u91cf\u5316-\">9. \u6a21\u578b\u5408\u5e76\u4e0e\u91cf\u5316 \ud83c\udd95<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p><strong>\u7ec4\u5408\u591a\u4e2a\u6a21\u578b\u6216\u538b\u7f29\u6a21\u578b<\/strong>\uff0c\u7528\u4e8e\u9ad8\u6548\u90e8\u7f72\u548c\u8bad\u7ec3\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"91-\u6a21\u578b\u5408\u5e76\">9.1 \u6a21\u578b\u5408\u5e76<\/h3>\n\n\n\n<h4 id=\"mergekit-\">MergeKit \u2b50\u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/arcee-ai\/mergekit\">https:\/\/github.com\/arcee-ai\/mergekit<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: Arcee AI<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u6700\u6d41\u884c\u7684 LLM \u6a21\u578b\u5408\u5e76\u5de5\u5177<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u652f\u6301 SLERP\u3001TIES\u3001DARE\u3001Passthrough\u3001\u8fdb\u5316\u5408\u5e76\u7b49\u65b9\u6cd5<\/li><li>Out-of-core \u65b9\u5f0f\uff0c\u4ec5\u9700 8GB VRAM \u5373\u53ef\u5728 CPU \u4e0a\u5408\u5e76<\/li><li>CMA-ES \u4f18\u5316\u5408\u5e76\u914d\u7f6e<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u514d\u8bad\u7ec3\u7684\u6a21\u578b\u80fd\u529b\u7ec4\u5408<\/li><\/ul>\n\n\n\n<h4 id=\"mergelm-\">MergeLM \u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/yule-BUAA\/MergeLM\">https:\/\/github.com\/yule-BUAA\/MergeLM<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: ICML 2024<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u7814\u7a76\u7ea7\u6a21\u578b\u5408\u5e76<\/li><\/ul>\n\n\n\n<h3 id=\"92-\u6a21\u578b\u91cf\u5316\">9.2 \u6a21\u578b\u91cf\u5316<\/h3>\n\n\n\n<h4 id=\"gptqmodel-\">GPTQModel \u2b50\u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/ModelCloud\/GPTQModel\">https:\/\/github.com\/ModelCloud\/GPTQModel<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u751f\u4ea7\u7ea7\u591a\u65b9\u6cd5\u91cf\u5316\u5de5\u5177<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u652f\u6301 GPTQ\u3001AWQ\u3001QQQ\u3001GPTAQ\u3001EoRA\u3001GAR<\/li><li>\u591a\u540e\u7aef\u652f\u6301\uff08CPU\/GPU\uff09<\/li><li>2026-03 \u652f\u6301 Qwen 3.5 MoE<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u751f\u4ea7\u90e8\u7f72\u91cf\u5316<\/li><\/ul>\n\n\n\n<h4 id=\"autogptq-\">AutoGPTQ \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/AutoGPTQ\/AutoGPTQ\">https:\/\/github.com\/AutoGPTQ\/AutoGPTQ<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u6613\u7528\u7684 GPTQ \u91cf\u5316<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>8\/4\/3\/2-bit \u91cf\u5316<\/li><li>Marlin int4*fp16 kernel<\/li><li>\u6708\u4e0b\u8f7d\u91cf ~150-200K<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5feb\u901f GPTQ \u91cf\u5316<\/li><\/ul>\n\n\n\n<h4 id=\"autoround-\">AutoRound \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/intel\/auto-round\">https:\/\/github.com\/intel\/auto-round<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: Intel<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u57fa\u4e8e\u7b26\u53f7\u68af\u5ea6\u4e0b\u964d\u7684\u9ad8\u7cbe\u5ea6\u91cf\u5316<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>2-4 bit \u9ad8\u7cbe\u5ea6<\/li><li>\u5bfc\u51fa\u683c\u5f0f\uff1aGPTQ\/AWQ\/GGUF\/AutoRound<\/li><li>\u5e7f\u6cdb\u786c\u4ef6\u517c\u5bb9\uff08CPU\u3001Intel GPU\u3001CUDA\u3001HPU\uff09<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u8de8\u5e73\u53f0\u90e8\u7f72\u91cf\u5316<\/li><\/ul>\n\n\n\n<h4 id=\"nvidia-model-optimizer--\">NVIDIA Model Optimizer \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/NVIDIA\/Model-Optimizer\">https:\/\/github.com\/NVIDIA\/Model-Optimizer<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u7edf\u4e00\u7684\u91cf\u5316\u3001\u526a\u679d\u3001\u84b8\u998f\u548c\u6295\u673a\u89e3\u7801\u4f18\u5316\u5e93<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>FP8\/INT8\/INT4 \u91cf\u5316 + \u7ed3\u6784\u5316\u526a\u679d + \u77e5\u8bc6\u84b8\u998f<\/li><li>\u5bfc\u51fa\u81f3 TensorRT-LLM\u3001TensorRT\u3001vLLM \u7b49\u90e8\u7f72\u6846\u67b6<\/li><li>NeMo Megatron Bridge \u652f\u6301 Nemotron-3-Super \u91cf\u5316\uff08PTQ \u548c QAT\uff09<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: NVIDIA \u751f\u6001\u7cfb\u7edf\u7684\u6a21\u578b\u538b\u7f29\u4e0e\u52a0\u901f<\/li><\/ul>\n\n\n\n<h4 id=\"turboquant-google--\">TurboQuant (Google) \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/research.google\/blog\/turboquant-redefining-ai-efficiency-with-extreme-compression\/\">https:\/\/research.google\/blog\/turboquant-redefining-ai-efficiency-with-extreme-compression\/<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: ICLR 2026<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: KV Cache \u6781\u9650\u538b\u7f29\uff0c\u96f6\u7cbe\u5ea6\u635f\u5931<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>3-bit \u91cf\u5316 KV Cache \u5b9e\u73b0 <strong>6x \u5185\u5b58\u7f29\u51cf<\/strong>\uff0c\u4e0b\u6e38\u4efb\u52a1\u5b8c\u7f8e\u4fdd\u6301<\/li><li>PolarQuant\uff08\u968f\u673a\u65cb\u8f6c\u7b80\u5316\u5411\u91cf\u51e0\u4f55\uff09+ QJL\uff081-bit \u6b8b\u5dee\u7ea0\u9519\uff09<\/li><li>H100 \u4e0a 4-bit \u914d\u7f6e\u8fbe <strong>8x \u6027\u80fd\u63d0\u5347<\/strong><\/li><li>\u65e0\u9700\u8bad\u7ec3\u6216\u5fae\u8c03<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u90e8\u7f72\u3001\u5185\u5b58\u53d7\u9650\u573a\u666f<\/li><\/ul>\n\n\n\n<h4 id=\"llamacpp-\">llama.cpp \u2b50\u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\">https:\/\/github.com\/ggml-org\/llama.cpp<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u6700\u901a\u7528\u7684 LLM \u63a8\u7406 + GGUF \u91cf\u5316<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>Q4_K_M \u6700\u4f73\u5e73\u8861\u70b9\uff1a92% \u8d28\u91cf\uff0c75% \u4f53\u79ef\u7f29\u51cf<\/li><li>CPU \/ GPU \/ Apple Silicon \u5168\u5e73\u53f0\u8fd0\u884c<\/li><li>ggml \u56e2\u961f\u5df2\u52a0\u5165 HuggingFace\uff082026-02\uff09<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u672c\u5730\u90e8\u7f72\u3001\u8fb9\u7f18\u8bbe\u5907<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"10-\u8f7b\u91cf\u7ea7\u9884\u8bad\u7ec3\u4e0e\u5206\u5e03\u5f0f\u8bad\u7ec3\">10. \u8f7b\u91cf\u7ea7\u9884\u8bad\u7ec3\u4e0e\u5206\u5e03\u5f0f\u8bad\u7ec3<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u914d\u5408 AutoResearch \u7b49\u81ea\u4e3b\u5b9e\u9a8c\u6846\u67b6\u4f7f\u7528\u2014\u2014<strong>\u5c0f\u89c4\u6a21\u5feb\u901f\u8bad\u7ec3\u662f\u81ea\u4e3b\u5b9e\u9a8c\u7684\u57fa\u7840<\/strong>\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"101-\u8f7b\u91cf\u7ea7\u9884\u8bad\u7ec3\">10.1 \u8f7b\u91cf\u7ea7\u9884\u8bad\u7ec3<\/h3>\n\n\n\n<h4 id=\"nanochat-karpathy-\">nanochat (Karpathy) \u2b50\u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/karpathy\/nanochat\">https:\/\/github.com\/karpathy\/nanochat<\/a><\/li><li><strong>\u5b9a\u4f4d<\/strong>: \u6700\u7b80 LLM \u8bad\u7ec3\u5b9e\u9a8c\u6846\u67b6\uff08AutoResearch \u7684\u5e95\u5c42\uff09<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u5355 GPU \u8282\u70b9\uff0c\u6781\u7b80\u4ee3\u7801<\/li><li>\u8986\u76d6\uff1aTokenization \u2192 \u9884\u8bad\u7ec3 \u2192 \u5fae\u8c03 \u2192 \u8bc4\u4f30 \u2192 \u63a8\u7406 \u2192 Chat UI<\/li><li><strong>\u7ea6 $48\uff088xH100 \u7ea6 2 \u5c0f\u65f6\uff09\u6216 $15\uff08Spot\uff09\u8bad\u7ec3 GPT-2 \u7ea7\u522b\u6a21\u578b<\/strong><\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5feb\u901f\u5b9e\u9a8c\u3001\u6559\u5b66\u3001AutoResearch \u5e95\u5c42<\/li><\/ul>\n\n\n\n<h4 id=\"nanotron-huggingface-\">Nanotron (HuggingFace) \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/huggingface\/nanotron\">https:\/\/github.com\/huggingface\/nanotron<\/a><\/li><li><strong>\u5b9a\u4f4d<\/strong>: \u6781\u7b80\u4f46\u652f\u6301 3D \u5e76\u884c\u7684 LLM \u9884\u8bad\u7ec3\u6846\u67b6<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u6570\u636e\u5e76\u884c + \u5f20\u91cf\u5e76\u884c + \u6d41\u6c34\u7ebf\u5e76\u884c<\/li><li>\u7075\u6d3b API\uff0c\u4f18\u5316\u5927\u89c4\u6a21\u8bad\u7ec3<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4ece\u5c0f\u5b9e\u9a8c\u5230\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u7684\u8fc7\u6e21<\/li><\/ul>\n\n\n\n<h3 id=\"102-\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6-\">10.2 \u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6 \ud83c\udd95<\/h3>\n\n\n\n<h4 id=\"torchtitan-\">TorchTitan \u2b50\u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/pytorch\/torchtitan\">https:\/\/github.com\/pytorch\/torchtitan<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: PyTorch \u5b98\u65b9<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: PyTorch \u539f\u751f\u7684\u5927\u89c4\u6a21\u8bad\u7ec3\u5e73\u53f0<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u6700\u9ad8 4D \u5e76\u884c\uff0c\u65e0\u9700\u4fee\u6539\u6a21\u578b\u4ee3\u7801<\/li><li>MXFP8 on Blackwell \u652f\u6301<\/li><li>262K token \u4e0a\u4e0b\u6587\u5e76\u884c<\/li><li>\u5f39\u6027\u6269\u7f29\u5bb9<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u548c\u5fae\u8c03<\/li><\/ul>\n\n\n\n<h4 id=\"open-dllm-\">Open-dLLM \u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/pengzhangzhi\/Open-dLLM\">https:\/\/github.com\/pengzhangzhi\/Open-dLLM<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u9996\u4e2a\u5f00\u6e90\u6269\u6563 LLM \u5168\u6808<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>: \u539f\u59cb\u6570\u636e \u2192 \u8bad\u7ec3 \u2192 Checkpoint \u2192 \u8bc4\u4f30 \u2192 \u63a8\u7406\uff0c\u4e00\u7ad9\u5f0f<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u6269\u6563\u8bed\u8a00\u6a21\u578b\u7814\u7a76<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"11-\u63a8\u7406\u5f15\u64cerl-\u8bad\u7ec3\u5faa\u73af\u6838\u5fc3\u7ec4\u4ef6\">11. \u63a8\u7406\u5f15\u64ce\uff08RL \u8bad\u7ec3\u5faa\u73af\u6838\u5fc3\u7ec4\u4ef6\uff09\ud83c\udd95<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u63a8\u7406\u5f15\u64ce\u5bf9 RL \u8bad\u7ec3\u81f3\u5173\u91cd\u8981\u2014\u2014<strong>80% \u7684 RLHF \u8bad\u7ec3\u65f6\u95f4\u7528\u4e8e\u751f\u6210\u6837\u672c<\/strong>\u3002\u5feb\u901f\u63a8\u7406 = \u5feb\u901f\u8bad\u7ec3\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"111-vllm-\">11.1 vLLM \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/vllm-project\/vllm\">https:\/\/github.com\/vllm-project\/vllm<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u6700\u6210\u719f\u7684\u5f00\u6e90 LLM \u63a8\u7406\u5f15\u64ce<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>PagedAttention \u5185\u5b58\u7ba1\u7406<\/li><li>v0.15.1: PyTorch 2.10\u3001Blackwell SM120 \u652f\u6301<\/li><li>Blackwell \u4e0a 4x \u541e\u5410\u91cf\u63d0\u5347<\/li><li>FP8\/FP4 \u6ce8\u610f\u529b\/GEMM kernel<\/li><li><strong>OpenRLHF \u7684\u6838\u5fc3\u63a8\u7406\u5f15\u64ce<\/strong><\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u751f\u4ea7\u63a8\u7406 + RL \u8bad\u7ec3\u91c7\u6837<\/li><\/ul>\n\n\n\n<h3 id=\"112-sglang-\">11.2 SGLang \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/sgl-project\/sglang\">https:\/\/github.com\/sgl-project\/sglang<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u9ad8\u6027\u80fd LLM &amp; \u591a\u6a21\u6001\u63a8\u7406<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>H100 \u4e0a\u7ea6 16,200 tok\/sec\uff08\u8d85\u8d8a vLLM \u7684 ~12,500\uff09<\/li><li>GB300 NVL72 \u4e0a 25x \u63a8\u7406\u6027\u80fd<\/li><li>RadixAttention \u524d\u7f00\u7f13\u5b58<\/li><li>\u539f\u751f TPU \u652f\u6301<\/li><li><strong>slime \u7684\u6838\u5fc3\u63a8\u7406\u5f15\u64ce<\/strong><\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9ad8\u541e\u5410\u63a8\u7406 + RL \u8bad\u7ec3\u91c7\u6837<\/li><\/ul>\n\n\n\n<h3 id=\"113-tensorrt-llm-\">11.3 TensorRT-LLM \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/NVIDIA\/TensorRT-LLM\">https:\/\/github.com\/NVIDIA\/TensorRT-LLM<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: NVIDIA<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u786c\u4ef6\u7ea7\u6df1\u5ea6\u4f18\u5316<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>FP8\/FP4\/INT4 AWQ\/INT8 SmoothQuant \u91cf\u5316<\/li><li>EAGLE-3 \u63a8\u6d4b\u89e3\u7801<\/li><li>KV Cache Connector API<\/li><li>AutoDeploy beta \u540e\u7aef<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u8ffd\u6c42\u6781\u81f4 GPU \u6027\u80fd<\/li><\/ul>\n\n\n\n<h3 id=\"114-lmdeploy-\">11.4 LMDeploy \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/InternLM\/lmdeploy\">https:\/\/github.com\/InternLM\/lmdeploy<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: LLM \u538b\u7f29\u3001\u90e8\u7f72\u3001\u63a8\u7406\u4e00\u4f53\u5316<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>TurboMind MXFP4 \u652f\u6301<\/li><li>H800 \u4e0a 1.5x vLLM \u6027\u80fd<\/li><li>DeepSeek PD \u89e3\u805a\u96c6\uff08DLSlime\/Mooncake \u96c6\u6210\uff09<\/li><li>\u53cc\u5f15\u64ce\u67b6\u6784\uff1aTurboMind (C++) + PyTorch (Python)<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u91cf\u5316\u6a21\u578b\u63a8\u7406\u670d\u52a1<\/li><\/ul>\n\n\n\n<h3 id=\"115-huggingface-tgi--\">11.5 HuggingFace TGI \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/huggingface\/text-generation-inference\">https:\/\/github.com\/huggingface\/text-generation-inference<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u591a\u540e\u7aef\u7edf\u4e00\u63a8\u7406\u524d\u7aef<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u652f\u6301\u591a\u540e\u7aef\uff1aTensorRT-LLM\u3001vLLM\u3001llama.cpp<\/li><li>Token \u6d41\u5f0f\u8f93\u51fa\u3001\u4f18\u5316\u6a21\u578b\u52a0\u8f7d\u3001\u5206\u5e03\u5f0f\u63a8\u7406<\/li><li>\u539f\u751f HF Hub \u96c6\u6210<\/li><li>CPU\/GPU\/Inferentia\/Trainium\/TPU \u652f\u6301<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: HuggingFace \u751f\u6001\u7684\u63a8\u7406\u670d\u52a1<\/li><\/ul>\n\n\n\n<h3 id=\"116-nvidia-dynamo-\">11.6 NVIDIA Dynamo \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/ai-dynamo\/dynamo\">https:\/\/github.com\/ai-dynamo\/dynamo<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: NVIDIA<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u6570\u636e\u4e2d\u5fc3\u7ea7\u5206\u5e03\u5f0f\u63a8\u7406<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>DeepSeek-R1 \u4e0a 30x \u8bf7\u6c42\u541e\u5410\u91cf<\/li><li>\u89e3\u805a\u96c6 prefill\/decode \u9636\u6bb5<\/li><li>\u52a8\u6001 GPU \u8c03\u5ea6 + LLM \u611f\u77e5\u8bf7\u6c42\u8def\u7531<\/li><li>Rust + Python \u6784\u5efa<\/li><li>\u652f\u6301 vLLM\u3001SGLang\u3001TensorRT-LLM \u4f5c\u4e3a\u540e\u7aef<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5927\u89c4\u6a21\u63a8\u7406\u96c6\u7fa4<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"12-\u591a\u6a21\u6001\u8bad\u7ec3\u6846\u67b6-\">12. \u591a\u6a21\u6001\u8bad\u7ec3\u6846\u67b6 \ud83c\udd95<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u8bad\u7ec3\u540c\u65f6\u7406\u89e3<strong>\u6587\u672c\u3001\u56fe\u50cf\u3001\u89c6\u9891\u3001\u97f3\u9891<\/strong>\u7684\u6a21\u578b\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"121-llava-onevision-15-\">12.1 LLaVA-OneVision-1.5 \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/EvolvingLMMs-Lab\/LLaVA-OneVision-1.5\">https:\/\/github.com\/EvolvingLMMs-Lab\/LLaVA-OneVision-1.5<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u5b8c\u5168\u5f00\u6e90\u7684\u591a\u6a21\u6001\u8bad\u7ec3\u6846\u67b6<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u539f\u751f\u5206\u8fa8\u7387\u56fe\u50cf\u5904\u7406<\/li><li>SOTA \u6027\u80fd + \u66f4\u4f4e\u8bad\u7ec3\u6210\u672c<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u591a\u6a21\u6001 LLM \u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"122-llava-onevision-15-rl-\">12.2 LLaVA-OneVision-1.5-RL \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/EvolvingLMMs-Lab\/LLaVA-OneVision-1.5-RL\">https:\/\/github.com\/EvolvingLMMs-Lab\/LLaVA-OneVision-1.5-RL<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u5f00\u6e90\u591a\u6a21\u6001 RL \u8bad\u7ec3<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>: \u4ee3\u7801 + \u6570\u636e + \u6a21\u578b\u5168\u5f00\u653e<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u591a\u6a21\u6001 RLHF<\/li><\/ul>\n\n\n\n<h3 id=\"123-openrlhf-m-\">12.3 OpenRLHF-M \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/OpenRLHF\/OpenRLHF-M\">https:\/\/github.com\/OpenRLHF\/OpenRLHF-M<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: OpenRLHF \u7684\u591a\u6a21\u6001\u6269\u5c55<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: VLM \u7684 RLHF \u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"124-llava-kd-\">12.4 LLaVA-KD \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Fantasyele\/LLaVA-KD\">https:\/\/github.com\/Fantasyele\/LLaVA-KD<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: ICCV 2025<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u591a\u6a21\u6001\u77e5\u8bc6\u84b8\u998f<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>: Multimodal Distillation + Relation Distillation<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5c06\u5927 MLLM \u84b8\u998f\u5230\u5c0f\u6a21\u578b<\/li><\/ul>\n\n\n\n<h3 id=\"125-moe-llava-\">12.5 MoE-LLaVA \u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/PKU-YuanGroup\/MoE-LLaVA\">https:\/\/github.com\/PKU-YuanGroup\/MoE-LLaVA<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: TMM 2025<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: MoE \u67b6\u6784\u7684\u89c6\u89c9\u8bed\u8a00\u6a21\u578b<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9ad8\u6548\u591a\u6a21\u6001 MoE \u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"126-llava-onevision-2--\">12.6 LLaVA-OneVision-2 \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/EvolvingLMMs-Lab\/LLaVA-OneVision-2\">https:\/\/github.com\/EvolvingLMMs-Lab\/LLaVA-OneVision-2<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2605.25979\uff08LLaVA-OneVision-2:\">https:\/\/arxiv.org\/abs\/2605.25979\uff08LLaVA-OneVision-2:<\/a> Towards Next-Generation Perceptual Intelligence\uff09<\/li><li><strong>\u9879\u76ee\u4e3b\u9875<\/strong>: <a href=\"https:\/\/evolvinglmms-lab.github.io\/LLaVA-OneVision-2\/\">https:\/\/evolvinglmms-lab.github.io\/LLaVA-OneVision-2\/<\/a><\/li><li><strong>\u53d1\u5e03<\/strong>: 2026-05-25<\/li><li><strong>\u6765\u6e90<\/strong>: EvolvingLMMs-Lab<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u4e0b\u4e00\u4ee3\u5b8c\u5168\u5f00\u6e90\u7684\u591a\u6a21\u6001\u8bad\u7ec3\u6846\u67b6\uff0c\u628a\u56fe\u50cf\u3001\u957f\u89c6\u9891\u3001\u7a7a\u95f4\u7406\u89e3\u7edf\u4e00\u5728\u5355\u4e00\u67b6\u6784\u4e0b<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u57fa\u4e8e\u539f\u751f <strong>OneVision-Encoder<\/strong>\uff08codec \u5bf9\u9f50\u89c6\u89c9\u7f16\u7801\u5668\uff09\uff0c\u56fe\u50cf\u3001\u5747\u5300\u91c7\u6837\u5e27\u3001codec \u5bf9\u9f50 token \u8d70\u540c\u4e00\u4f4d\u7f6e\u7f16\u7801\u65b9\u6848\uff0c\u65e0\u9700\u4e3a\u4e0d\u540c\u8f93\u5165\u7c7b\u578b\u914d\u7f6e\u4e0d\u540c tokenizer<\/li><li>Windowed Attention \u5728\u4fdd\u6301\u539f\u751f\u5206\u8fa8\u7387\u7684\u540c\u65f6\u5b9e\u73b0\u9ad8\u6548\u5c40\u90e8\u8ba1\u7b97<\/li><li>8B \u7ea7\u6a21\u578b\uff0c\u5355\u4e00\u6a21\u578b\u3001\u539f\u751f\u5206\u8fa8\u7387\u3001\u65e0\u9700\u4efb\u52a1\u4e13\u7528 adapter<\/li><li><strong>\u7aef\u5230\u7aef\u5168\u5f00\u653e\u4e14\u53ef\u590d\u73b0<\/strong>\uff1a\u6570\u636e\u3001\u7f16\u7801\u5668\u3001\u8bad\u7ec3\u3001checkpoint\u3001\u65e5\u5fd7\u9010\u9636\u6bb5\u5168\u90e8\u91ca\u51fa<\/li><li>\u56db\u6bb5\u5f0f\u7d27\u51d1\u8bad\u7ec3\uff1a\u4ece LLaVA-OneVision-1.5 \u5f15\u5bfc\u89c6\u9891\u80fd\u529b\uff0830s \u77ed\u89c6\u9891\u5b57\u5e55\uff09\u2192 \u5927\u89c4\u6a21\u591a\u6a21\u6001\u6307\u4ee4\u5fae\u8c03\uff0830\u2013180s \u89c6\u9891\u5b57\u5e55\uff09<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u56fe\u50cf+\u957f\u89c6\u9891+\u7a7a\u95f4\u7edf\u4e00\u7406\u89e3\u7684\u591a\u6a21\u6001 LLM \u8bad\u7ec3\uff1b\u8ffd\u6c42\u5b8c\u6574\u53ef\u590d\u73b0\u8bad\u7ec3\u914d\u65b9<\/li><\/ul>\n\n\n\n<h3 id=\"127-verl-omni--\">12.7 VeRL-Omni \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/verl-project\/verl-omni\">https:\/\/github.com\/verl-project\/verl-omni<\/a><\/li><li><strong>\u5b98\u65b9\u535a\u5ba2<\/strong>: <a href=\"https:\/\/vllm.ai\/blog\/2026-05-14-verl-omni\">https:\/\/vllm.ai\/blog\/2026-05-14-verl-omni<\/a><\/li><li><strong>\u6587\u6863<\/strong>: <a href=\"https:\/\/verl-omni.readthedocs.io\/\">https:\/\/verl-omni.readthedocs.io\/<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: verl \u9879\u76ee\u5b98\u65b9\uff08\u4ece verl \u7684\u591a\u6a21\u6001\u751f\u6210 RL \u5206\u652f\u72ec\u7acb\uff09<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u9762\u5411<strong>\u591a\u6a21\u6001\u751f\u6210\u6a21\u578b<\/strong>\u7684\u901a\u7528 RL \u8bad\u7ec3\u6846\u67b6\u2014\u2014&#8221;Easy, Fast, and Stable RL Training for Diffusion and Omni-Modality Models&#8221;<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u8986\u76d6\u4e09\u7c7b\u751f\u6210\u6a21\u578b<\/strong>\uff1a\u56fe\u50cf\/\u89c6\u9891\/\u97f3\u9891\u6269\u6563\u6a21\u578b\uff08Qwen-Image\u3001Wan2.2\u3001LTX-2.3\uff09\u3001\u7edf\u4e00\u7406\u89e3+\u751f\u6210\u6a21\u578b\uff08BAGEL\u3001HunyuanImage-3.0\uff09\u3001\u5168\u6a21\u6001\u6a21\u578b\uff08Qwen3-Omni\uff09<\/li><li><strong>vLLM-Omni \u4e13\u7528 rollout<\/strong>\uff1arequest \u7ea7 \/ step-wise \u6279\u5904\u7406 + FA3\uff0c\u9ad8\u541e\u5410\u6269\u6563\u4e0e\u591a\u6a21\u6001\u751f\u6210\u91c7\u6837<\/li><li><strong>\u7075\u6d3b\u7684\u5956\u52b1\u6d41\u6c34\u7ebf<\/strong>\uff1a\u89c4\u5219\u5956\u52b1\u3001\u6a21\u578b\u5956\u52b1\u3001\u591a\u6a21\u6001\u5956\u52b1\u8ba1\u7b97\u53ef\u81ea\u7531\u7ec4\u5408<\/li><li><strong>\u6a21\u5757\u5316\u8bad\u7ec3\u540e\u7aef<\/strong>\uff1a\u590d\u7528\u5df2\u6709\u5e76\u884c\u7b56\u7565\uff08FSDP\u3001USP\uff09\u4e0e\u4f18\u5316\u624b\u6bb5\uff0c\u800c\u975e\u4ece\u96f6\u91cd\u5efa\u8bad\u7ec3\u6808<\/li><li>\u5df2\u96c6\u6210 DiffusionNFT\u3001Diffusion DPO\uff08Qwen-Image \/ SD3.5 \u5df2\u9a8c\u8bc1\u914d\u65b9\uff09\uff1bv0.2.0 \u91cd\u5efa Qwen3-Omni \u591a\u6a21\u6001\u8bad\u7ec3\uff08DPO &amp; GSPO\uff09<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u56fe\u50cf\/\u89c6\u9891\/\u97f3\u9891\u751f\u6210\u6a21\u578b\u7684 RL \u540e\u8bad\u7ec3\uff1b\u5168\u6a21\u6001\u6a21\u578b\u5bf9\u9f50<\/li><\/ul>\n\n\n\n<h3 id=\"128-verl-vla--\">12.8 verl-vla \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/verl-project\/verl-vla\">https:\/\/github.com\/verl-project\/verl-vla<\/a><\/li><li><strong>\u53d1\u5e03<\/strong>: v0.1.0 \u2014 2026-08<\/li><li><strong>\u6765\u6e90<\/strong>: verl \u9879\u76ee\u5b98\u65b9<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: \u7edf\u4e00\u7684<strong>\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c\uff08VLA\uff09<\/strong>\u7b56\u7565\u540e\u8bad\u7ec3\u6846\u67b6\u2014\u2014\u628a\u4eba\u5728\u56de\u8def\u6570\u636e\u91c7\u96c6\u3001\u76d1\u7763\u5fae\u8c03\u3001\u5f3a\u5316\u5b66\u4e60\u3001\u7b56\u7565\u8bc4\u6d4b\u6536\u8fdb\u540c\u4e00\u5957\u5de5\u4f5c\u6d41<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u6a21\u578b\u3001\u73af\u5883\u3001\u8bad\u7ec3\u7b97\u6cd5\u53ef\u72ec\u7acb\u63a5\u5165<\/strong>\uff0c\u901a\u8fc7\u5171\u4eab\u6267\u884c\u67b6\u6784\u81ea\u7531\u7ec4\u5408<\/li><li><strong>\u8de8\u4e91\u8fb9\u5206\u5e03\u5f0f\u90e8\u7f72<\/strong>\uff1a\u8bad\u7ec3 Worker\u3001\u4eff\u771f\u5668\u3001\u5b9e\u4f53\u673a\u5668\u4eba\u53ef\u6309\u5404\u81ea\u7684\u786c\u4ef6\u4e0e\u8fde\u901a\u6027\u9700\u6c42\u8dd1\u5728\u4e0d\u540c\u8282\u70b9\u4e0a<\/li><li><strong>\u4eba\u5728\u56de\u8def<\/strong>\uff1a\u64cd\u4f5c\u8005\u53ef\u4ece\u4efb\u610f\u8054\u7f51\u8bbe\u5907\u76d1\u63a7\u3001\u9065\u64cd\u4f5c\u3001\u4ecb\u5165\u4eff\u771f\u73af\u5883\u4e0e\u771f\u673a\u7684\u6267\u884c\u8fc7\u7a0b<\/li><li>\u57fa\u4e8e verl \u6784\u5efa\uff0c\u590d\u7528\u5176\u6210\u719f\u7684 RL \u540e\u8bad\u7ec3\u6808<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u5177\u8eab\u667a\u80fd \/ \u673a\u5668\u4eba\u7b56\u7565\u7684 RL \u540e\u8bad\u7ec3\uff1b\u9700\u8981\u771f\u673a + \u4eff\u771f\u6df7\u5408\u8bad\u7ec3\u4e0e\u4eba\u5de5\u4ecb\u5165\u7684\u573a\u666f<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"13-\u5b9e\u9a8c\u8ffd\u8e2a\u4e0e\u7f16\u6392\u5e73\u53f0\">13. \u5b9e\u9a8c\u8ffd\u8e2a\u4e0e\u7f16\u6392\u5e73\u53f0<\/h2>\n\n\n\n<h3 id=\"131-weights--biases-wb-\">13.1 Weights &amp; Biases (W&amp;B) \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/wandb.ai\">https:\/\/wandb.ai<\/a><\/li><li><strong>\u529f\u80fd<\/strong>: \u5b9e\u9a8c\u8ffd\u8e2a + Sweeps \u8d85\u53c2\u6570\u4f18\u5316 + \u6a21\u578b\u7248\u672c\u7ba1\u7406 + \u62a5\u544a<\/li><li><strong>\u7279\u70b9<\/strong>: \u884c\u4e1a\u6807\u51c6\uff0c\u4e0e\u6240\u6709\u4e3b\u6d41\u8bad\u7ec3\u6846\u67b6\u96c6\u6210<\/li><\/ul>\n\n\n\n<h3 id=\"132-mlflow-30-\">13.2 MLflow 3.0 \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u529f\u80fd<\/strong>: \u5b9e\u9a8c\u8ffd\u8e2a + \u8d85\u53c2\u6570\u626b\u63cf + \u6a21\u578b\u6ce8\u518c + \u6a21\u578b\u670d\u52a1<\/li><li><strong>\u7279\u70b9<\/strong>: \u5f00\u6e90\u3001\u81ea\u6258\u7ba1\u3001\u652f\u6301\u5d4c\u5957\u5b9e\u9a8c<\/li><\/ul>\n\n\n\n<h3 id=\"133-clearml--\">13.3 ClearML \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/allegroai\/clearml\">https:\/\/github.com\/allegroai\/clearml<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u5f00\u6e90 MLOps \u5e73\u53f0<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>150K+ \u7528\u6237\u904d\u5e03 Fortune 500 \u4f01\u4e1a<\/li><li>\u81ea\u52a8\u65e5\u5fd7\u8bb0\u5f55<\/li><li>Pipeline \u7f16\u6392<\/li><li>\u6a21\u578b\u670d\u52a1<\/li><li>\u6570\u636e\u96c6\u7248\u672c\u7ba1\u7406<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u4f01\u4e1a\u7ea7 MLOps<\/li><\/ul>\n\n\n\n<h3 id=\"134-huggingface-trackio-\">13.4 HuggingFace Trackio \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>\u5b9a\u4f4d<\/strong>: HF \u751f\u6001\u7684\u8f7b\u91cf\u7ea7\u5b9e\u9a8c\u8ffd\u8e2a<\/li><li><strong>\u7279\u70b9<\/strong>: \u4e0e HF Skills \u6df1\u5ea6\u96c6\u6210\uff0cAgent \u53ef\u76f4\u63a5\u8bfb\u53d6\u6307\u6807\u5e76\u51b3\u7b56<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"14-benchmark-\u4e0e\u8bc4\u6d4b\u4f53\u7cfb\">14. Benchmark \u4e0e\u8bc4\u6d4b\u4f53\u7cfb<\/h2>\n\n\n\n<h3 id=\"141-ml-agent-\u8bc4\u6d4b\u57fa\u51c6\">14.1 ML Agent \u8bc4\u6d4b\u57fa\u51c6<\/h3>\n\n\n\n<h4 id=\"mle-bench-\">MLE-bench \u2b50\u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/arxiv.org\/abs\/2410.07095\">https:\/\/arxiv.org\/abs\/2410.07095<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: OpenAI<\/li><li><strong>\u5185\u5bb9<\/strong>: 75 \u4e2a Kaggle ML \u5de5\u7a0b\u7ade\u8d5b\u4efb\u52a1<\/li><li><strong>\u8bc4\u4f30<\/strong>: AI Agent \u8bad\u7ec3\u6a21\u578b\u3001\u51c6\u5907\u6570\u636e\u3001\u8fd0\u884c\u5b9e\u9a8c\u7684\u80fd\u529b<\/li><\/ul>\n\n\n\n<h4 id=\"mlagentbench-\">MLAgentBench \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/snap-stanford\/MLAgentBench\">https:\/\/github.com\/snap-stanford\/MLAgentBench<\/a><\/li><li><strong>\u8bc4\u4f30<\/strong>: AI Agent \u7aef\u5230\u7aef ML \u5b9e\u9a8c\u80fd\u529b\uff0813 \u4e2a\u4efb\u52a1\uff09<\/li><\/ul>\n\n\n\n<h4 id=\"paperbench--\">PaperBench \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/openai.com\/index\/paperbench\/\">https:\/\/openai.com\/index\/paperbench\/<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: OpenAI<\/li><li><strong>\u8bc4\u4f30<\/strong>: AI \u4ece\u96f6\u590d\u73b0 ICML 2024 \u8bba\u6587\u7684\u80fd\u529b<\/li><li><strong>\u5173\u952e\u6570\u636e<\/strong>: 8,316 \u4e2a\u53ef\u8bc4\u5206\u4efb\u52a1\uff0c\u8986\u76d6 20 \u7bc7\u8bba\u6587\uff1b\u6700\u4f73 Agent \u5f97\u5206 21%<\/li><\/ul>\n\n\n\n<h4 id=\"core-bench--\">CORE-Bench \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/openreview.net\/forum?id=BsMMc4MEGS\">https:\/\/openreview.net\/forum?id=BsMMc4MEGS<\/a><\/li><li><strong>\u8bc4\u4f30<\/strong>: \u8ba1\u7b97\u53ef\u590d\u73b0\u6027 Agent \u8bc4\u6d4b<\/li><li><strong>\u5173\u952e\u6570\u636e<\/strong>: 90 \u7bc7\u8bba\u6587\u7684 270 \u4e2a\u4efb\u52a1\uff0c\u8de8 CS\u3001\u793e\u4f1a\u79d1\u5b66\u3001\u533b\u5b66<\/li><\/ul>\n\n\n\n<h4 id=\"mlrc-bench-\">MLRC-Bench \u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/openreview.net\/forum?id=t8Okk2PRWU\">https:\/\/openreview.net\/forum?id=t8Okk2PRWU<\/a><\/li><li><strong>\u8bc4\u4f30<\/strong>: AI Agent \u89e3\u51b3 ML \u7814\u7a76\u7ade\u8d5b\u7684\u80fd\u529b<\/li><\/ul>\n\n\n\n<h4 id=\"agentbench--\">AgentBench \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/THUDM\/AgentBench\">https:\/\/github.com\/THUDM\/AgentBench<\/a><\/li><li><strong>\u8bc4\u4f30<\/strong>: LLM Agent \u591a\u7ef4\u5ea6\u8bc4\u6d4b<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>: \u8986\u76d6 OS\u3001\u6570\u636e\u5e93\u3001\u77e5\u8bc6\u56fe\u8c31\u3001Web\u3001\u6e38\u620f\u73af\u5883<\/li><\/ul>\n\n\n\n<h4 id=\"swe-bench-verified--\">SWE-bench Verified \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/www.swebench.com\/\">https:\/\/www.swebench.com\/<\/a><\/li><li><strong>\u8bc4\u4f30<\/strong>: \u4eba\u5de5\u9a8c\u8bc1\u7684 GitHub Issue \u89e3\u51b3<\/li><li><strong>\u5173\u952e\u6570\u636e<\/strong>: \u7f16\u7801 Agent \u884c\u4e1a\u6807\u51c6\uff1b\u9876\u7ea7\u5f97\u5206 70%+<\/li><\/ul>\n\n\n\n<h4 id=\"livebench--\">LiveBench \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/livebench.ai\/\">https:\/\/livebench.ai\/<\/a><\/li><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/LiveBench\/LiveBench\">https:\/\/github.com\/LiveBench\/LiveBench<\/a><\/li><li><strong>\u8bc4\u4f30<\/strong>: \u6297\u6c61\u67d3\u3001\u65e0\u9700 LLM \u88c1\u5224\u7684\u7efc\u5408 LLM \u8bc4\u6d4b<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>6 \u5927\u7c7b\u522b\uff1a\u6570\u5b66\u3001\u63a8\u7406\u3001\u6570\u636e\u5206\u6790\u3001\u8bed\u8a00\u3001\u7f16\u7a0b\u3001\u6307\u4ee4\u9075\u4ece<\/li><li>\u9898\u76ee\u6bcf\u6708\u66f4\u65b0\uff0c\u57fa\u4e8e\u6700\u65b0\u6570\u5b66\u7ade\u8d5b\u3001arXiv \u8bba\u6587\u3001\u65b0\u95fb<\/li><li>\u6240\u6709\u9898\u76ee\u6709\u5ba2\u89c2\u53ef\u9a8c\u8bc1\u7b54\u6848\uff0c\u81ea\u52a8\u8bc4\u5206\uff0c\u65e0\u9700 LLM Judge<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u65e0\u6c61\u67d3\u98ce\u9669\u7684 LLM \u80fd\u529b\u8bc4\u6d4b<\/li><\/ul>\n\n\n\n<h4 id=\"agenticdatabench--\">AgenticDataBench \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/huggingface.co\/papers\/2607.01647\uff08arXiv\">https:\/\/huggingface.co\/papers\/2607.01647\uff08arXiv<\/a> 2607.01647\uff09<\/li><li><strong>\u53d1\u5e03<\/strong>: 2026-07-02<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u9762\u5411 LLM \u6570\u636e Agent\uff08Data Agent\uff09\u7684\u7efc\u5408\u8bc4\u6d4b\u57fa\u51c6<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u8986\u76d6 <strong>15 \u4e2a\u9886\u57df<\/strong> + 5 \u4e2a\u771f\u5b9e B2B \u91d1\u878d\u79d1\u6280\u7528\u4f8b\uff0c\u8d34\u8fd1\u771f\u5b9e\u6570\u636e\u79d1\u5b66\u5de5\u4f5c\u6d41<\/li><li>\u63d0\u4f9b<strong>\u7ec6\u7c92\u5ea6\u6280\u80fd\u7ea7\u6807\u7b7e<\/strong>\uff1aschema \u68c0\u67e5\u3001\u8868\u8fde\u63a5\u3001\u6570\u636e\u6e05\u6d17\u3001\u53ef\u89c6\u5316\u3001\u4e1a\u52a1\u4e0a\u4e0b\u6587\u63a8\u7406\u2014\u2014\u800c\u975e\u53ea\u770b\u5355\u4e00\u4efb\u52a1\u6210\u529f\u7387<\/li><li>\u9644\u5e26 GitHub \u6d4b\u8bd5\u53f0 + Hugging Face \u6570\u636e\u96c6\uff0cApache-2.0\uff0c\u6613\u4e8e\u68c0\u67e5\u4e0e\u590d\u73b0<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u8bc4\u4f30\u81ea\u52a8\u5316\u6570\u636e\u5904\u7406\/\u6570\u636e\u79d1\u5b66 Agent \u662f\u5426\u5177\u5907\u771f\u5b9e\u5de5\u4f5c\u6d41\u80fd\u529b<\/li><\/ul>\n\n\n\n<h4 id=\"terminal-bench-21--\">Terminal-Bench 2.1 \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/laude-institute\/terminal-bench-2\">https:\/\/github.com\/laude-institute\/terminal-bench-2<\/a><\/li><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/www.tbench.ai\/\">https:\/\/www.tbench.ai\/<\/a><\/li><li><strong>\u8bba\u6587<\/strong>: <a href=\"https:\/\/huggingface.co\/papers\/2601.11868\">https:\/\/huggingface.co\/papers\/2601.11868<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: \u65af\u5766\u798f\u5927\u5b66 + Laude Institute + Terminal-Bench \u5f00\u6e90\u793e\u533a<\/li><li><strong>\u8bc4\u4f30<\/strong>: Agent \u5728\u547d\u4ee4\u884c\u754c\u9762\u4e2d\u5b8c\u6210\u786c\u6838\u771f\u5b9e\u4efb\u52a1\u7684\u80fd\u529b\u2014\u2014<strong>89 \u4e2a\u7cbe\u9009\u4efb\u52a1<\/strong><\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u4efb\u52a1\u8986\u76d6\u8f6f\u4ef6\u5de5\u7a0b\u3001\u7cfb\u7edf\u7ba1\u7406\u3001\u6570\u636e\u5904\u7406\u3001\u5b89\u5168\u3001\u79d1\u5b66\u8ba1\u7b97\uff0c\u4ee5\u53ca<strong>\u6a21\u578b\u8bad\u7ec3<\/strong>\u2014\u2014\u76f4\u63a5\u5bf9\u5e94\u672c\u62a5\u544a\u5173\u5fc3\u7684\u81ea\u52a8\u5316\u8bad\u7ec3\u573a\u666f<\/li><li>\u6bcf\u4e2a\u4efb\u52a1\u7ed9 Agent \u4e00\u6bb5\u4e66\u9762\u6307\u4ee4 + \u4e00\u4e2a\u542b\u76f8\u5173\u6587\u4ef6\/\u8f6f\u4ef6\/\u5de5\u5177\u7684\u4e13\u7528\u5bb9\u5668\uff1bAgent \u53ef\u68c0\u89c6\u73af\u5883\u3001\u6267\u884c\u547d\u4ee4\u3001\u7f16\u8f91\u6587\u4ef6\u3001\u5b89\u88c5\u4f9d\u8d56\u5e76\u6839\u636e\u6267\u884c\u53cd\u9988\u4fee\u6b63\u65b9\u6848\uff0c\u6700\u540e\u7531\u81ea\u52a8\u5316\u6d4b\u8bd5\u68c0\u67e5\u5bb9\u5668\u7ec8\u6001<\/li><li>\u91c7\u7528 <strong>Harbor \u4efb\u52a1\u683c\u5f0f\u4e0e Harness<\/strong> \u8fd0\u884c\uff0c\u5f00\u7bb1\u652f\u6301 Claude Code\u3001Codex CLI\u3001OpenHands\u3001Mini-SWE-Agent\u3001Terminus 2<\/li><li><strong>v2.1 \u4fee\u590d\u4e86 2.0 \u4e2d 89 \u4e2a\u4efb\u52a1\u91cc\u7684 28 \u4e2a<\/strong>\uff08\u5916\u90e8\u4f9d\u8d56\u53d8\u66f4\u3001\u8d44\u6e90\u9884\u7b97\u8fc7\u7d27\u3001\u6307\u4ee4\u4e0e\u6d4b\u8bd5\u4e0d\u5339\u914d\uff09\uff0c\u4fee\u590d\u540e\u4e0d\u5b58\u5728&#8221;\u65e0\u4eba\u53ef\u89e3&#8221;\u7684\u4efb\u52a1\u2014\u2014\u5931\u8d25\u771f\u5b9e\u53cd\u6620 Agent \u80fd\u529b\u800c\u975e\u73af\u5883\u7f3a\u9677<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u6a2a\u5411\u5bf9\u6bd4\u4e0d\u540c\u7f16\u7801 Agent \/ Harness \u7684\u771f\u5b9e\u7ec8\u7aef\u4f5c\u4e1a\u80fd\u529b\uff1b\u8bc4\u4f30 Agent \u80fd\u5426\u80dc\u4efb\u8bad\u7ec3\u811a\u672c\u7f16\u5199\u4e0e\u8fd0\u7ef4\u7c7b\u4efb\u52a1<\/li><\/ul>\n\n\n\n<h3 id=\"142-\u6a21\u578b\u8bc4\u6d4b\u6846\u67b6-\">14.2 \u6a21\u578b\u8bc4\u6d4b\u6846\u67b6 \ud83c\udd95<\/h3>\n\n\n\n<h4 id=\"deepeval-\">DeepEval \u2b50\u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/confident-ai\/deepeval\">https:\/\/github.com\/confident-ai\/deepeval<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u7c7b Pytest \u7684 LLM \u8bc4\u6d4b\u6846\u67b6<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>v3.0: \u7ec4\u4ef6\u7ea7\u7c92\u5ea6<\/li><li>14+ \u8bc4\u6d4b\u6307\u6807\uff08G-Eval\u3001\u5e7b\u89c9\u68c0\u6d4b\u7b49\uff09<\/li><li>\u591a\u8f6e\u5bf9\u8bdd\u6a21\u62df<\/li><li>DeepTeam \u7ea2\u961f\u6d4b\u8bd5<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: CI\/CD \u4e2d\u96c6\u6210 LLM \u8bc4\u6d4b<\/li><\/ul>\n\n\n\n<h4 id=\"opik-\">Opik \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/comet-ml\/opik\">https:\/\/github.com\/comet-ml\/opik<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: Comet<\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u5f00\u6e90 LLM \u53ef\u89c2\u6d4b\u6027\u4e0e\u8bc4\u6d4b<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>\u6df1\u5ea6\u8ffd\u8e2a<\/li><li>LLM-as-a-Judge \u6307\u6807<\/li><li>\u5e7b\u89c9\u68c0\u6d4b<\/li><li>\u751f\u4ea7\u73af\u5883\u4eea\u8868\u76d8<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u751f\u4ea7\u73af\u5883 LLM \u76d1\u63a7<\/li><\/ul>\n\n\n\n<h4 id=\"lmms-eval-\">LMMs-Eval \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/EvolvingLMMs-Lab\/lmms-eval\">https:\/\/github.com\/EvolvingLMMs-Lab\/lmms-eval<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u8de8\u6587\u672c\u3001\u56fe\u50cf\u3001\u89c6\u9891\u3001\u97f3\u9891\u7684\u591a\u6a21\u6001\u8bc4\u6d4b<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li>v0.6: eval-as-a-service HTTP \u670d\u52a1\u5668<\/li><li>7.5x \u541e\u5410\u91cf\u63d0\u5347<\/li><li>50+ \u65b0\u4efb\u52a1<\/li><li>\u7edf\u8ba1\u4e25\u683c\u6027\uff08\u7f6e\u4fe1\u533a\u95f4\u3001\u914d\u5bf9 t \u68c0\u9a8c\uff09<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u591a\u6a21\u6001\u6a21\u578b\u5168\u9762\u8bc4\u6d4b<\/li><\/ul>\n\n\n\n<h4 id=\"arize-phoenix-\">Arize Phoenix \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Arize-ai\/phoenix\">https:\/\/github.com\/Arize-ai\/phoenix<\/a><\/li><li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>: \u5f00\u6e90 LLM \u53ef\u89c2\u6d4b\u6027\u4e0e\u8bc4\u6d4b<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>: \u5b8c\u5168\u81ea\u6258\u7ba1\uff1b\u8ffd\u8e2a\u3001\u8bc4\u6d4b\u3001\u68c0\u7d22\u5206\u6790<\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u81ea\u90e8\u7f72\u7684 LLM \u76d1\u63a7<\/li><\/ul>\n\n\n\n<h4 id=\"livecodebench-\">LiveCodeBench \u2b50\u2b50\u2b50\u2b50<\/h4>\n\n\n\n<ul><li><strong>\u6392\u884c\u699c<\/strong>: <a href=\"https:\/\/artificialanalysis.ai\/evaluations\/livecodebench\">https:\/\/artificialanalysis.ai\/evaluations\/livecodebench<\/a><\/li><li><strong>\u8bc4\u4f30<\/strong>: LLM \u4ee3\u7801\u751f\u6210\u80fd\u529b\uff08\u65e0\u6570\u636e\u6c61\u67d3\uff09<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"15-\u8f85\u52a9\u7f16\u7801-agent\u53ef\u7528\u4e8e\u8bad\u7ec3\u811a\u672c\u5f00\u53d1\">15. \u8f85\u52a9\u7f16\u7801 Agent\uff08\u53ef\u7528\u4e8e\u8bad\u7ec3\u811a\u672c\u5f00\u53d1\uff09<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u8fd9\u4e9b Agent \u672c\u8eab\u4e0d\u76f4\u63a5\u8bad\u7ec3\u6a21\u578b\uff0c\u4f46\u53ef\u4ee5<strong>\u7f16\u5199\u548c\u8c03\u8bd5\u8bad\u7ec3\u4ee3\u7801<\/strong>\uff0c\u914d\u5408 HF Skills \u7b49\u5b9e\u73b0\u5b8c\u6574\u81ea\u52a8\u5316\u3002<\/p><\/blockquote>\n\n\n\n<h3 id=\"151-aider-\">15.1 Aider \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/Aider-AI\/aider\">https:\/\/github.com\/Aider-AI\/aider<\/a><\/li><li><strong>\u7528\u9014<\/strong>: \u7ec8\u7aef AI \u914d\u5bf9\u7f16\u7a0b\uff0c\u7f16\u5199\/\u4fee\u6539\u8bad\u7ec3\u811a\u672c<\/li><li><strong>\u4f18\u52bf<\/strong>: Git \u96c6\u6210\u3001\u652f\u6301\u591a\u79cd LLM<\/li><\/ul>\n\n\n\n<h3 id=\"152-openhands-\">15.2 OpenHands \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/OpenHands\/OpenHands\">https:\/\/github.com\/OpenHands\/OpenHands<\/a><\/li><li><strong>\u7528\u9014<\/strong>: \u81ea\u4e3b\u7f16\u5199\u4ee3\u7801 + \u6267\u884c + \u8c03\u8bd5<\/li><li><strong>\u4f18\u52bf<\/strong>: \u53ef\u81ea\u52a8\u4fee\u6539\u8bad\u7ec3\u811a\u672c\u5e76\u8fd0\u884c<\/li><\/ul>\n\n\n\n<h3 id=\"153-swe-agent-\">15.3 SWE-agent \u2b50\u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/SWE-agent\/SWE-agent\">https:\/\/github.com\/SWE-agent\/SWE-agent<\/a><\/li><li><strong>\u7528\u9014<\/strong>: \u81ea\u52a8\u4fee\u590d\u8bad\u7ec3\u4ee3\u7801\u4e2d\u7684 Bug<\/li><li><strong>\u4f18\u52bf<\/strong>: SWE-bench \u5f00\u6e90 SOTA<\/li><\/ul>\n\n\n\n<h3 id=\"154-open-swe--\">15.4 Open-SWE \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/langchain-ai\/open-swe\">https:\/\/github.com\/langchain-ai\/open-swe<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: LangChain<\/li><li><strong>\u7528\u9014<\/strong>: \u5f02\u6b65\u4e91\u7aef\u7f16\u7801 Agent<\/li><li><strong>\u4f18\u52bf<\/strong>: \u591a Agent \u67b6\u6784\uff08Planner + Reviewer\uff09\uff1bGitHub \u96c6\u6210\uff1b\u81ea\u52a8\u521b\u5efa PR<\/li><\/ul>\n\n\n\n<h3 id=\"155-sera--\">15.5 SERA \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>\u6765\u6e90<\/strong>: AllenAI<\/li><li><strong>GitHub<\/strong>: <a href=\"https:\/\/huggingface.co\/collections\/allenai\/open-coding-agents\">https:\/\/huggingface.co\/collections\/allenai\/open-coding-agents<\/a><\/li><li><strong>\u7528\u9014<\/strong>: \u5f00\u6e90\u7f16\u7801 Agent \u5bb6\u65cf<\/li><li><strong>\u4f18\u52bf<\/strong>: SWE-Bench \u4e0a 54.2%\uff1b\u4ec5 40 GPU-days (~$2K) \u8bad\u7ec3\uff1b\u4ee3\u7801\/\u6570\u636e\/\u6a21\u578b\u5168\u5f00\u653e<\/li><\/ul>\n\n\n\n<h3 id=\"156-cline--\">15.6 Cline \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/cline\/cline\">https:\/\/github.com\/cline\/cline<\/a><\/li><li><strong>\u7528\u9014<\/strong>: VS Code \u5185\u81ea\u4e3b\u7f16\u7801 Agent<\/li><li><strong>\u4f18\u52bf<\/strong>: 60K+ GitHub stars\uff1b5M+ \u5f00\u53d1\u8005\u4fe1\u4efb\uff1bMCP \u5de5\u5177\u6269\u5c55\uff1b\u539f\u751f subagents (v3.58)\uff1bhuman-in-the-loop \u5ba1\u6279\u673a\u5236<\/li><\/ul>\n\n\n\n<h3 id=\"157-opencode--\">15.7 OpenCode \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/opencode-ai\/opencode\">https:\/\/github.com\/opencode-ai\/opencode<\/a><\/li><li><strong>\u7528\u9014<\/strong>: Go \u8bed\u8a00\u6784\u5efa\u7684\u7ec8\u7aef AI \u7f16\u7801 Agent<\/li><li><strong>\u4f18\u52bf<\/strong>: 95K+ GitHub stars\uff1bBubble Tea TUI \u4ea4\u4e92\u754c\u9762\uff1b75+ LLM \u63d0\u4f9b\u5546\uff1b650 \u4e07\u6708\u6d3b\u5f00\u53d1\u8005\uff1bSQLite \u6301\u4e45\u5316<\/li><\/ul>\n\n\n\n<h3 id=\"158-plandex--\">15.8 Plandex \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/plandex-ai\/plandex\">https:\/\/github.com\/plandex-ai\/plandex<\/a><\/li><li><strong>\u7528\u9014<\/strong>: \u9762\u5411\u5927\u578b\u9879\u76ee\u7684\u7ec8\u7aef\u7f16\u7801 Agent<\/li><li><strong>\u4f18\u52bf<\/strong>: 2M token \u4e0a\u4e0b\u6587\uff1btree-sitter \u9879\u76ee\u6620\u5c04\uff0830+ \u8bed\u8a00\uff09\uff1bdiff review \u6c99\u76d2\uff1b\u81ea\u52a8\u8c03\u8bd5\u6d4f\u89c8\u5668\u5e94\u7528<\/li><\/ul>\n\n\n\n<h3 id=\"159-roo-code--\">15.9 Roo Code \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/RooVetGit\/Roo-Code\">https:\/\/github.com\/RooVetGit\/Roo-Code<\/a><\/li><li><strong>\u7528\u9014<\/strong>: \u7ec8\u7aef\u7f16\u7801 Agent<\/li><li><strong>\u4f18\u52bf<\/strong>: 95K+ GitHub stars\uff1b75+ LLM \u63d0\u4f9b\u5546\uff1b250 \u4e07\u6708\u6d3b\u5f00\u53d1\u8005<\/li><\/ul>\n\n\n\n<h3 id=\"1510-gemini-cli-\">15.10 Gemini CLI \u2b50\u2b50\u2b50\u2b50<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/google-gemini\/gemini-cli\">https:\/\/github.com\/google-gemini\/gemini-cli<\/a><\/li><li><strong>\u6765\u6e90<\/strong>: Google\uff08Apache 2.0\uff09<\/li><li><strong>\u7528\u9014<\/strong>: Google \u5b98\u65b9\u5f00\u6e90\u7ec8\u7aef AI Agent\uff0c\u76f4\u63a5\u5728\u547d\u4ee4\u884c\u8c03\u7528 Gemini \u6a21\u578b\u8fdb\u884c\u7f16\u7801\u3001\u8c03\u8bd5\u4e0e\u4efb\u52a1\u81ea\u52a8\u5316<\/li><li><strong>\u4f18\u52bf<\/strong>:<ul><li><strong>\u514d\u8d39\u989d\u5ea6\u5145\u88d5<\/strong>\uff1a\u4e2a\u4eba Google \u8d26\u53f7\u6bcf\u5206\u949f 60 \u6b21\u3001\u6bcf\u5929 1000 \u6b21\u8bf7\u6c42\uff0c\u53ef\u514d\u8d39\u4f7f\u7528 Gemini 2.5 Pro<\/li><li>\u652f\u6301 MCP Server \u6269\u5c55\uff0c\u53ef\u63a5\u5165\u5916\u90e8\u5de5\u5177\u548c\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd<\/li><li>\u6bcf\u5468\u53d1\u5e03\u7a33\u5b9a\u7248\uff0c\u6bcf\u5468\u53d1\u5e03\u9884\u89c8\u7248\uff0c\u8fed\u4ee3\u6d3b\u8dc3<\/li><li>\u4e0e HF Skills \u573a\u666f B \u7ec4\u5408\u2014\u2014&#8221;\u7528 Gemini CLI \u9a71\u52a8\u8bad\u7ec3\u6d41\u7a0b&#8221;<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: Google \u751f\u6001\u7528\u6237\uff1b\u514d\u8d39\u8d44\u6e90\u53d7\u9650\u573a\u666f\uff1b\u914d\u5408 HF Skills \u505a\u81ea\u7136\u8bed\u8a00\u9a71\u52a8\u8bad\u7ec3<\/li><\/ul>\n\n\n\n<h3 id=\"1511-claw-code--\">15.11 Claw Code \u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/ultraworkers\/claw-code\">https:\/\/github.com\/ultraworkers\/claw-code<\/a><\/li><li><strong>\u5b98\u7f51<\/strong>: <a href=\"https:\/\/claw-code.codes\/\">https:\/\/claw-code.codes\/<\/a><\/li><li><strong>\u53d1\u5e03<\/strong>: 2026-04-02<\/li><li><strong>\u6765\u6e90<\/strong>: Sigrid Jin \/ ultraworkers \u793e\u533a<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: Claude Code \u67b6\u6784\u7684\u5f00\u6e90 Rust \u91cd\u5199\u7248\u2014\u2014\u4e3a\u5f00\u53d1\u8005\u63d0\u4f9b\u5b8c\u5168\u900f\u660e\u3001\u53ef\u5b9a\u5236\u3001\u53ef\u81ea\u6258\u7ba1\u7684\u7ec8\u7aef AI \u7f16\u7801 Agent<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>100K+ GitHub stars<\/strong>\uff0cGitHub \u53f2\u4e0a\u6700\u5feb\u7a81\u7834 10 \u4e07 star \u7684\u9879\u76ee\u4e4b\u4e00<\/li><li>Clean-Room \u91cd\u5199\uff08Rust ~4K \u884c + Python ~1.5K \u884c\uff09\uff0c\u65e0\u4e13\u6709\u4ee3\u7801\uff0c\u5b8c\u5168\u53ef\u5ba1\u8ba1<\/li><li><strong>\u591a\u63d0\u4f9b\u5546\u652f\u6301<\/strong>\uff1aAnthropic\u3001OpenAI\u3001xAI\/Grok\u3001Alibaba Qwen\u3001Ollama\u3001OpenRouter \u7b49\u4efb\u610f OpenAI \u517c\u5bb9\u7aef\u70b9<\/li><li>\u6838\u5fc3\u5de5\u5177\u96c6\u5b8c\u6574\uff1aBash\u3001Read\u3001Write\u3001Edit\u3001Glob\u3001Grep\u2014\u2014\u4e0e Claude Code \u5de5\u5177\u5c42\u5bf9\u7b49<\/li><li>\u4f1a\u8bdd\u6301\u4e45\u5316\uff08.claude\/sessions \u76ee\u5f55\uff09+ OAuth \u8ba4\u8bc1\u6d41<\/li><li>\u652f\u6301\u672c\u5730\u6a21\u578b\uff08Ollama\/LM Studio\uff09\uff0c\u65e0\u9700\u4e91\u7aef API<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u5b8c\u5168\u5f00\u6e90\u3001\u53ef\u79bb\u7ebf\u8fd0\u884c\u7684 Agent \u73af\u5883\uff1b\u914d\u5408 HF Skills \u505a\u81ea\u7136\u8bed\u8a00\u9a71\u52a8\u8bad\u7ec3\uff1b\u5bf9 Claude Code \u67b6\u6784\u611f\u5174\u8da3\u7684\u7814\u7a76\u8005<\/li><\/ul>\n\n\n\n<h3 id=\"1512-deepseek-harness-dsh--\">15.12 DeepSeek Harness (dsh) \u2b50\u2b50\u2b50\u2b50\u2b50 \ud83c\udd95<\/h3>\n\n\n\n<ul><li><strong>GitHub<\/strong>: <a href=\"https:\/\/github.com\/deepseek-ai\/deepseek-harness\">https:\/\/github.com\/deepseek-ai\/deepseek-harness<\/a><\/li><li><strong>\u53d1\u5e03<\/strong>: 2026-08-13<\/li><li><strong>\u6765\u6e90<\/strong>: DeepSeek AI \u5b98\u65b9<\/li><li><strong>\u6838\u5fc3\u7406\u5ff5<\/strong>: &#8220;Everything is a Plugin&#8221;\u2014\u2014DeepSeek \u5b98\u65b9\u5f00\u6e90\u7684 Agent Harness\uff0c\u628a Agent \u8fd0\u884c\u65f6\u7684\u6bcf\u4e00\u5c42\u90fd\u505a\u6210\u53ef\u66ff\u6362\u63d2\u4ef6<\/li><li><strong>\u5173\u952e\u7279\u6027<\/strong>:<ul><li><strong>\u4e24\u5929\u7a81\u7834 10 \u4e07 GitHub star<\/strong>\uff0c\u4e24\u5468\u7ea6 20 \u4e07 star\uff0c\u662f GitHub \u6709\u8bb0\u5f55\u4ee5\u6765\u6700\u5feb\u7684\u5f00\u53d1\u8005\u5de5\u5177\u91c7\u7eb3\u66f2\u7ebf\u4e4b\u4e00\uff1bMIT \u8bb8\u53ef<\/li><li><strong>\u4e00\u5207\u7686\u63d2\u4ef6<\/strong>\uff1a\u6a21\u578b\u9002\u914d\u5668\u3001\u5de5\u5177\u6ce8\u518c\u8868\u3001\u4f1a\u8bdd\u65e5\u5fd7\u3001\u6c99\u7bb1\uff0c\u4e43\u81f3 <strong>Agent \u4e3b\u5faa\u73af\u672c\u8eab<\/strong>\u90fd\u662f\u63d2\u4ef6\uff0c\u7531 Cordis \u63d2\u4ef6\u6846\u67b6\u9a71\u52a8<\/li><li>Node.js \u5b9e\u73b0\uff0c\u5b8c\u5168\u672c\u5730\u8fd0\u884c\uff1b\u4e0e Claude Code\u3001OpenAI Codex \u5bf9\u6807\u540c\u4e00\u7247\u9886\u5730<\/li><li>Web-UI \u4f18\u5148\uff08\u800c\u975e TUI\uff09\uff1b\u76ee\u524d\u662f developer preview\uff0c\u660e\u786e\u4f1a\u6709 breaking change<\/li><\/ul><\/li><li><strong>\u9002\u7528\u573a\u666f<\/strong>: \u9700\u8981\u6df1\u5ea6\u5b9a\u5236 Agent \u5faa\u73af\/\u5de5\u5177\u5c42\u6765\u9a71\u52a8\u8bad\u7ec3\u6d41\u7a0b\u7684\u56e2\u961f\uff1b\u81ea\u5efa\u8bad\u7ec3 Agent \u57fa\u7840\u8bbe\u65bd<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"16-\u603b\u7ed3\u5bf9\u6bd4\u8868\u4e0e\u9009\u578b\u6307\u5357\">16. \u603b\u7ed3\u5bf9\u6bd4\u8868\u4e0e\u9009\u578b\u6307\u5357<\/h2>\n\n\n\n<h3 id=\"\u6838\u5fc3\u5de5\u5177\u5bf9\u6bd4\">\u6838\u5fc3\u5de5\u5177\u5bf9\u6bd4<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u5de5\u5177<\/th><th>\u5b9a\u4f4d<\/th><th>\u81ea\u52a8\u5316\u5c42\u7ea7<\/th><th>\u5f00\u6e90<\/th><th>\u6838\u5fc3\u4ef7\u503c<\/th><\/tr><\/thead><tbody><tr><td><strong>AutoResearch<\/strong><\/td><td>\u81ea\u4e3b\u5b9e\u9a8c\u5faa\u73af<\/td><td>\ud83d\udd34 \u5168\u81ea\u52a8<\/td><td>\u2705 MIT<\/td><td>AI \u81ea\u4e3b\u505a ML \u5b9e\u9a8c<\/td><\/tr><tr><td><strong>AI Scientist v2<\/strong><\/td><td>\u81ea\u4e3b\u79d1\u5b66\u7814\u7a76<\/td><td>\ud83d\udd34 \u5168\u81ea\u52a8<\/td><td>\u2705<\/td><td>AI \u81ea\u4e3b\u505a\u7814\u7a76 + \u5199\u8bba\u6587<\/td><\/tr><tr><td><strong>AutoML-Agent<\/strong><\/td><td>\u5168\u6d41\u7a0b AutoML<\/td><td>\ud83d\udd34 \u5168\u81ea\u52a8<\/td><td>\u2705<\/td><td>\u591a Agent \u534f\u4f5c AutoML<\/td><\/tr><tr><td><strong>HF Skills<\/strong><\/td><td>Agent \u8bad\u7ec3\u80fd\u529b<\/td><td>\ud83d\udfe1 \u81ea\u7136\u8bed\u8a00\u9a71\u52a8<\/td><td>\u2705<\/td><td>\u4e00\u53e5\u8bdd\u5b8c\u6210\u6a21\u578b\u8bad\u7ec3\u5168\u6d41\u7a0b<\/td><\/tr><tr><td><strong>HF AutoTrain<\/strong><\/td><td>\u65e0\u4ee3\u7801\u8bad\u7ec3<\/td><td>\ud83d\udfe1 \u5411\u5bfc\u5f0f<\/td><td>\u2705<\/td><td>\u4e0a\u4f20\u6570\u636e\u81ea\u52a8\u8bad\u7ec3<\/td><\/tr><tr><td><strong>Unsloth<\/strong><\/td><td>\u9ad8\u6548\u5fae\u8c03\u5f15\u64ce<\/td><td>\ud83d\udfe2 \u9700\u914d\u7f6e<\/td><td>\u2705<\/td><td>2x \u901f\u5ea6 70% \u7701\u663e\u5b58<\/td><\/tr><tr><td><strong>LlamaFactory<\/strong><\/td><td>Web UI \u5fae\u8c03<\/td><td>\ud83d\udfe1 \u53ef\u89c6\u5316<\/td><td>\u2705<\/td><td>\u6d4f\u89c8\u5668\u4e2d\u914d\u7f6e\u8bad\u7ec3<\/td><\/tr><tr><td><strong>Axolotl<\/strong><\/td><td>\u751f\u4ea7\u5fae\u8c03\u6846\u67b6<\/td><td>\ud83d\udfe2 YAML \u914d\u7f6e<\/td><td>\u2705<\/td><td>\u7075\u6d3b + \u751f\u4ea7\u5c31\u7eea<\/td><\/tr><tr><td><strong>OpenRLHF<\/strong><\/td><td>RLHF \u6846\u67b6<\/td><td>\ud83d\udfe2 \u9700\u914d\u7f6e<\/td><td>\u2705<\/td><td>70B+ \u5168\u91cf RLHF<\/td><\/tr><tr><td><strong>verl<\/strong><\/td><td>RL \u8bad\u7ec3\u6846\u67b6<\/td><td>\ud83d\udfe2 \u9700\u914d\u7f6e<\/td><td>\u2705<\/td><td>\u5de5\u4e1a\u7ea7 RL\uff08\u5b57\u8282\/\u963f\u91cc\/Berkeley\uff09<\/td><\/tr><tr><td><strong>DAPO<\/strong><\/td><td>RL \u7b97\u6cd5<\/td><td>\ud83d\udfe2 \u9700\u914d\u7f6e<\/td><td>\u2705<\/td><td>SOTA \u63a8\u7406 RL \u8bad\u7ec3<\/td><\/tr><tr><td><strong>AReaL<\/strong><\/td><td>\u5f02\u6b65 RL \u8bad\u7ec3<\/td><td>\ud83d\udfe2 \u9700\u914d\u7f6e<\/td><td>\u2705<\/td><td>2.77x \u52a0\u901f<\/td><\/tr><tr><td><strong>slime<\/strong><\/td><td>RL \u540e\u8bad\u7ec3\u6846\u67b6<\/td><td>\ud83d\udfe2 \u9700\u914d\u7f6e<\/td><td>\u2705<\/td><td>\u9a71\u52a8 GLM \u7cfb\u5217\u6a21\u578b<\/td><\/tr><tr><td><strong>rLLM<\/strong><\/td><td>Agent RL \u8bad\u7ec3<\/td><td>\ud83d\udfe2 \u9700\u914d\u7f6e<\/td><td>\u2705<\/td><td>Agent \u80fd\u529b RL \u8bad\u7ec3<\/td><\/tr><tr><td><strong>Distilabel<\/strong><\/td><td>\u5408\u6210\u6570\u636e\u751f\u6210<\/td><td>\ud83d\udfe1 Pipeline \u5f0f<\/td><td>\u2705<\/td><td>\u89c4\u6a21\u5316 SFT\/DPO \u6570\u636e<\/td><\/tr><tr><td><strong>Magpie<\/strong><\/td><td>\u5bf9\u9f50\u6570\u636e\u5408\u6210<\/td><td>\ud83d\udd34 \u5168\u81ea\u52a8<\/td><td>\u2705<\/td><td>\u96f6 prompt \u5de5\u7a0b<\/td><\/tr><tr><td><strong>EasyDistill<\/strong><\/td><td>\u77e5\u8bc6\u84b8\u998f<\/td><td>\ud83d\udfe1 \u534a\u81ea\u52a8<\/td><td>\u2705<\/td><td>\u9ed1\u76d2 + \u767d\u76d2\u84b8\u998f<\/td><\/tr><tr><td><strong>MergeKit<\/strong><\/td><td>\u6a21\u578b\u5408\u5e76<\/td><td>\ud83d\udfe1 \u914d\u7f6e\u5f0f<\/td><td>\u2705<\/td><td>\u514d\u8bad\u7ec3\u80fd\u529b\u7ec4\u5408<\/td><\/tr><tr><td><strong>SPIN<\/strong><\/td><td>\u81ea\u8fdb\u5316\u8bad\u7ec3<\/td><td>\ud83d\udfe1 \u534a\u81ea\u52a8<\/td><td>\u2705<\/td><td>\u65e0\u9700\u989d\u5916\u504f\u597d\u6570\u636e<\/td><\/tr><tr><td><strong>Optuna<\/strong><\/td><td>\u8d85\u53c2\u6570\u4f18\u5316<\/td><td>\ud83d\udfe1 \u81ea\u52a8\u641c\u7d22<\/td><td>\u2705<\/td><td>\u884c\u4e1a\u6807\u51c6 HPO<\/td><\/tr><tr><td><strong>NNI<\/strong><\/td><td>AutoML \u5168\u6808<\/td><td>\ud83d\udfe1 \u81ea\u52a8\u641c\u7d22<\/td><td>\u2705<\/td><td>NAS + HPO + \u6a21\u578b\u538b\u7f29<\/td><\/tr><tr><td><strong>nanochat<\/strong><\/td><td>\u8f7b\u91cf\u9884\u8bad\u7ec3<\/td><td>\ud83d\udfe2 \u624b\u52a8<\/td><td>\u2705<\/td><td>AutoResearch \u5e95\u5c42<\/td><\/tr><tr><td><strong>DeepEval<\/strong><\/td><td>\u6a21\u578b\u8bc4\u6d4b<\/td><td>\ud83d\udfe1 \u81ea\u52a8\u5316<\/td><td>\u2705<\/td><td>CI\/CD \u96c6\u6210\u8bc4\u6d4b<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 id=\"\u6309\u573a\u666f\u9009\u578b\">\u6309\u573a\u666f\u9009\u578b<\/h3>\n\n\n\n<h4 id=\"\u573a\u666f-a\u8ba9-ai-\u5168\u81ea\u52a8\u8dd1\u5b9e\u9a8c\u81ea\u4e3b\u4f18\u5316\u6a21\u578b\">\u573a\u666f A\uff1a\u8ba9 AI \u5168\u81ea\u52a8\u8dd1\u5b9e\u9a8c\u3001\u81ea\u4e3b\u4f18\u5316\u6a21\u578b<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>AutoResearch + nanochat + Optuna\nAI Scientist v2\uff08\u66f4\u504f\u7814\u7a76\u4fa7\uff09\nAutoML-Agent\uff08ICML 2025\uff0c\u591a Agent \u534f\u4f5c\uff09\nArbor\uff08\u5047\u8bbe\u6811\u7d2f\u79ef\u7ecf\u9a8c\uff0c\u540c\u9884\u7b97\u4e0b 2.5x \u4e8e Codex \/ Claude Code\uff09\nSibyl-AutoResearch\uff08\u81ea\u8fdb\u5316\u8bd5\u9519 Harness\uff0c20+ Agent \u7814\u7a76\u7ec4\u7ec7\uff09<\/code><\/pre>\n\n\n\n<h4 id=\"\u573a\u666f-b\u7528\u81ea\u7136\u8bed\u8a00\u6307\u6325-agent-\u5b8c\u6210\u6a21\u578b\u8bad\u7ec3\">\u573a\u666f B\uff1a\u7528\u81ea\u7136\u8bed\u8a00\u6307\u6325 Agent \u5b8c\u6210\u6a21\u578b\u8bad\u7ec3<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>HF Skills\uff08model-trainer + jobs + trackio + evaluation\uff09\n+ Claude Code \/ Codex \/ Gemini CLI<\/code><\/pre>\n\n\n\n<h4 id=\"\u573a\u666f-c\u9ad8\u6548\u5fae\u8c03\u73b0\u6709\u6a21\u578b\">\u573a\u666f C\uff1a\u9ad8\u6548\u5fae\u8c03\u73b0\u6709\u6a21\u578b<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>\u5355 GPU: Unsloth\uff08\u901f\u5ea6\u738b\uff09\n\u591a GPU: Axolotl \/ LlamaFactory\uff08\u7075\u6d3b + \u6613\u7528\uff09\nNVIDIA \u96c6\u7fa4: NeMo AutoModel\nPyTorch \u539f\u6559\u65e8: torchtune\n\u65e0\u4ee3\u7801: H2O LLM Studio \/ HF AutoTrain<\/code><\/pre>\n\n\n\n<h4 id=\"\u573a\u666f-drlhf--\u5bf9\u9f50\u8bad\u7ec3\">\u573a\u666f D\uff1aRLHF \/ \u5bf9\u9f50\u8bad\u7ec3<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>\u6807\u51c6 RLHF: OpenRLHF\n\u5b57\u8282\u751f\u6001: verl + DAPO\n\u5f02\u6b65\u52a0\u901f: AReaL\uff082.77x \u52a0\u901f\uff09\nGLM \u751f\u6001: slime\nNVIDIA \u751f\u6001: NeMo RL + NeMo Gym\nAgent RL: rLLM \/ RAGEN \/ OpenManus-RL\n\u7b80\u5355\u63a8\u7406 RL: SimpleRL-Reason\nGRPO\uff08\u63a8\u7406\u8bad\u7ec3\uff09: TRL \/ Unsloth \/ Axolotl \u5747\u652f\u6301\nf-\u6563\u5ea6 GRPO: f-GRPO\uff08\u66f4\u7075\u6d3b\u7684\u7b56\u7565\u4f18\u5316\uff09\nAgent \u6811\u641c\u7d22 RL: Tree-GRPO\uff084x \u91c7\u6837\u6548\u7387\uff09\nRL \u7cfb\u7edf\u7ea7\u63d0\u901f: Arctic RL\uff08ZoRRo \u96f6\u5197\u4f59 Rollout\uff0c3.5x \u7aef\u5230\u7aef\u52a0\u901f\uff09\nK8s \u81ea\u6258\u7ba1\u540e\u8bad\u7ec3: OpenRL\uff08Google\uff0cTinker \u517c\u5bb9 API\uff09\n\u5728\u7ebf\/\u5e38\u9a7b\u5f0f RL: OpenClaw-RL\uff08\u90e8\u7f72\u5bf9\u8bdd\u5373\u8bad\u7ec3\u4fe1\u53f7\uff09\nvLLM \u751f\u6001\u539f\u751f RL: vime\uff08Megatron \u8bad\u7ec3 + vLLM rollout\uff0c\u8bad\u63a8\u5bf9\u9f50\u7a33\u5b9a\uff09\n\u6269\u6563 LLM \u7684 RL: dLLM-RL \/ TraceRL\uff08\u9996\u4e2a diffusion LLM \u540e\u8bad\u7ec3\u6846\u67b6\uff09\n\u4f01\u4e1a\u7ea7\u5927\u89c4\u6a21 MoE RL: Miles\uff08SGLang rollout + Megatron \u8bad\u7ec3\uff0c\u4e07\u4ebf\u53c2\u6570\u7ea7\uff09\n\u8de8\u73af\u5883\u901a\u7528 Agent RL: Open-AgentRL\uff08RLAnything \/ AutoTool \/ DemyAgent\uff09\n\u6269\u6563\u4e0e\u5168\u6a21\u6001\u751f\u6210 RL: VeRL-Omni\uff08Qwen-Image \/ Wan2.2 \/ Qwen3-Omni\uff09\n\u5177\u8eab\u667a\u80fd VLA \u540e\u8bad\u7ec3: verl-vla\uff08\u4eba\u5728\u56de\u8def + \u4eff\u771f\/\u771f\u673a\u6df7\u5408\uff09<\/code><\/pre>\n\n\n\n<h4 id=\"\u573a\u666f-e\u81ea\u8fdb\u5316--\u51cf\u5c11\u4eba\u5de5\u6807\u6ce8\">\u573a\u666f E\uff1a\u81ea\u8fdb\u5316 \/ \u51cf\u5c11\u4eba\u5de5\u6807\u6ce8<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>SPIN \u2192 \u81ea\u5bf9\u5f08\u5fae\u8c03\nSPC \u2192 \u5bf9\u6297\u5f0f\u63a8\u7406\u8bc4\u5224\u8fdb\u5316\nSPELL \u2192 \u957f\u4e0a\u4e0b\u6587\u81ea\u5bf9\u5f08\nMulti-Agent Evolve \u2192 \u591a\u89d2\u8272\u534f\u540c\u8fdb\u5316\nMultiagent Finetuning \u2192 \u591a Agent \u6301\u7eed\u8fed\u4ee3<\/code><\/pre>\n\n\n\n<h4 id=\"\u573a\u666f-f\u5408\u6210\u6570\u636e--\u8bad\u7ec3--\u8bc4\u6d4b\">\u573a\u666f F\uff1a\u5408\u6210\u6570\u636e \u2192 \u8bad\u7ec3 \u2192 \u8bc4\u6d4b<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>Distilabel \/ Magpie \u2192 \u6570\u636e\u751f\u6210\nNeMo Curator \/ DataTrove \u2192 \u6570\u636e\u6e05\u6d17\nUnsloth \/ TRL \u2192 \u8bad\u7ec3\nDeepEval \/ LMMs-Eval \u2192 \u8bc4\u6d4b<\/code><\/pre>\n\n\n\n<h4 id=\"\u573a\u666f-g\u6a21\u578b\u538b\u7f29\u4e0e\u90e8\u7f72\">\u573a\u666f G\uff1a\u6a21\u578b\u538b\u7f29\u4e0e\u90e8\u7f72<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>\u77e5\u8bc6\u84b8\u998f: EasyDistill \/ DistillKit\n\u6a21\u578b\u5408\u5e76: MergeKit\n\u91cf\u5316: GPTQModel \/ llama.cpp GGUF<\/code><\/pre>\n\n\n\n<h4 id=\"\u573a\u666f-h\u65e0\u4ee3\u7801\u5feb\u901f\u539f\u578b\">\u573a\u666f H\uff1a\u65e0\u4ee3\u7801\u5feb\u901f\u539f\u578b<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>HF AutoTrain\uff08\u4e0a\u4f20\u6570\u636e\u5373\u53ef\uff09\nLlamaFactory LlamaBoard\uff08Web UI\uff09\nH2O LLM Studio\uff08\u6d4f\u89c8\u5668 GUI\uff09<\/code><\/pre>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"\u8d8b\u52bf\u89c2\u5bdf2026-q3-\u66f4\u65b0\">\u8d8b\u52bf\u89c2\u5bdf\uff082026 Q3 \u66f4\u65b0\uff09<\/h2>\n\n\n\n<ol><li><strong>AutoResearch \u8303\u5f0f\u5174\u8d77<\/strong>: Karpathy \u8bc1\u660e 630 \u884c\u4ee3\u7801\u5373\u53ef\u5b9e\u73b0&#8221;AI \u81ea\u4e3b\u505a ML \u7814\u7a76&#8221;\uff082026-04 \u5df2 66K+ stars\uff09\u2014\u2014\u73b0\u5df2\u5f62\u6210\u5b8c\u6574\u884d\u751f\u751f\u6001\uff1aARIS\u3001AI-Supervisor\u3001AutoResearchClaw\uff08\u60f3\u6cd5\u2192\u8bba\u6587\uff09\u3001Deli_AutoResearch\uff08\u957f\u5468\u671f\u534f\u8bae\uff09\u3001\u9886\u57df\u65e0\u5173\u79fb\u690d\u7248 pi-autoresearch<\/li><li><strong>&#8220;Vibe Training&#8221; \u5230\u6765<\/strong>: HF Skills \u8ba9\u7528\u81ea\u7136\u8bed\u8a00\u9a71\u52a8\u5b8c\u6574\u8bad\u7ec3\u6d41\u7a0b\u6210\u4e3a\u73b0\u5b9e<\/li><li><strong>GRPO \u53d8\u4f53\u7206\u53d1<\/strong>: f-GRPO\uff08f-\u6563\u5ea6\u5bb6\u65cf\uff09\u3001Tree-GRPO\uff08\u6811\u641c\u7d22\uff0cICLR 2026\uff09\u3001DAPO\u2014\u2014GRPO \u5df2\u6210\u9ed8\u8ba4\u5bf9\u9f50\u65b9\u6cd5\uff0c\u4e13\u95e8\u53d8\u4f53\u5feb\u901f\u6d8c\u73b0<\/li><li><strong>RL \u6846\u67b6\u767e\u82b1\u9f50\u653e<\/strong>: verl\u3001DAPO\u3001AReaL\u3001slime\u3001NeMo RL\u2014\u2014\u5404\u5927\u5382\u548c\u5b9e\u9a8c\u5ba4\u7eb7\u7eb7\u5f00\u6e90 RL \u8bad\u7ec3\u6846\u67b6<\/li><li><strong>Self-Play \u7a81\u7834\u74f6\u9888<\/strong>: \u591a Agent \u81ea\u8fdb\u5316\u8bad\u7ec3\uff08SPIN\u3001MAE\u3001SPC\u3001SPELL\uff09\u6b63\u5728\u89e3\u51b3\u5355\u6a21\u578b\u81ea\u8bad\u7ec3\u7684 plateau \u95ee\u9898<\/li><li><strong>\u5408\u6210\u6570\u636e\u6210\u4e3a\u57fa\u7840\u8bbe\u65bd<\/strong>: Distilabel\u3001Magpie\u3001Evidently \u4f7f\u6570\u636e\u751f\u6210\u6210\u4e3a\u8bad\u7ec3 Pipeline \u7684\u4e00\u7b49\u516c\u6c11\uff1b\u6a21\u578b\u574d\u7f29\u5bf9\u7b56\uff08Evol-Instruct\uff09\u6210\u4e3a\u6807\u914d<\/li><li><strong>MCP \u534f\u8bae\u6807\u51c6\u5316<\/strong>: Agent \u8fde\u63a5\u8bad\u7ec3\u5de5\u5177\u7684\u6807\u51c6\u534f\u8bae\uff0c\u88ab OpenAI \/ Google \/ Microsoft \u91c7\u7eb3<\/li><li><strong>\u5355 GPU \u4e5f\u80fd\u505a\u7814\u7a76<\/strong>: Unsloth + nanochat + AutoResearch \u7ec4\u5408\u8ba9\u4e2a\u4eba\u5f00\u53d1\u8005\u4e5f\u80fd\u9ad8\u6548\u505a LLM \u5b9e\u9a8c<\/li><li><strong>\u63a8\u7406-\u8bad\u7ec3\u4e00\u4f53\u5316<\/strong>: vLLM\/SGLang\/TGI \u5df2\u6210\u4e3a RL \u8bad\u7ec3\u5faa\u73af\u7684\u6838\u5fc3\u7ec4\u4ef6\uff0c\u4e0d\u4ec5\u4ec5\u662f\u63a8\u7406\u670d\u52a1<\/li><li><strong>\u591a\u6a21\u6001 RL<\/strong>: LLaVA-OneVision-1.5-RL \u548c OpenRLHF-M \u5c06 RL \u5bf9\u9f50\u6269\u5c55\u5230\u89c6\u89c9\u8bed\u8a00\u6a21\u578b<\/li><li><strong>\u6781\u9650\u91cf\u5316\u65f6\u4ee3<\/strong>: Google TurboQuant\uff08ICLR 2026\uff09\u5b9e\u73b0 3-bit KV Cache \u96f6\u7cbe\u5ea6\u635f\u5931 6x \u538b\u7f29\uff1bNVIDIA Model Optimizer \u7edf\u4e00\u91cf\u5316\/\u526a\u679d\/\u84b8\u998f<\/li><li><strong>\u591a Agent \u7f16\u7801\u6d6a\u6f6e<\/strong>: 2026 \u5e74 2 \u6708\u5404\u5927\u5de5\u5177\u96c6\u4e2d\u53d1\u5e03\u591a Agent \u80fd\u529b\uff08Grok Build\u3001Windsurf\u3001Claude Code\u3001Codex CLI\u3001Devin\uff09\u2014\u2014\u7f16\u7801 Agent \u73b0\u5728\u5e38\u89c4\u7f16\u5199\u8bad\u7ec3\u811a\u672c<\/li><li><strong>Kernel \u7ea7 Autoresearch<\/strong>: 2026-04 \u65b0\u8d8b\u52bf\u2014\u2014Meta KernelAgent\u3001RightNow AI AutoKernel \u628a &#8220;\u8ba9 AI \u7761\u89c9\u65f6\u505a\u7814\u7a76&#8221; \u7684\u8303\u5f0f\u4e0b\u6c89\u5230 GPU kernel \u751f\u6210\uff0c\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\u672c\u8eab\u4e5f\u5f00\u59cb\u88ab Agent \u81ea\u52a8\u4f18\u5316<\/li><li><strong>\u591a Agent RL \u7cfb\u7edf<\/strong>: MARTI\u3001rLLM\u3001ART\u3001OpenManus-RL\u2014\u2014RL \u8bad\u7ec3\u5bf9\u8c61\u6b63\u4ece\u5355\u4e2a\u6a21\u578b\u6269\u5c55\u5230\u591a Agent \u7cfb\u7edf\uff1b\u4e2d\u5fc3\u5316\u591a Agent \u4ea4\u4e92 + \u5206\u5e03\u5f0f\u7b56\u7565\u8bad\u7ec3\u6210\u4e3a\u65b0\u8303\u5f0f<\/li><li><strong>Agent \u9a71\u52a8\u6570\u636e\u521b\u9020<\/strong>: Meta synthetic-data-kit\uff08\u6587\u6863 \u2192 \u6570\u636e\u96c6\u7684 4 \u6b65 CLI\uff09\u548c Autodata \u6846\u67b6\u663e\u793a\u2014\u2014\u8ba9 AI Agent \u8fed\u4ee3\u4f18\u5316\u81ea\u5df1\u7684\u6570\u636e\u751f\u6210\u8fc7\u7a0b\uff0c\u6548\u679c\u663e\u8457\u4f18\u4e8e\u6807\u51c6\u5408\u6210\u6570\u636e\u65b9\u6cd5<\/li><li><strong>Autoresearch \u6210\u4e3a\u901a\u7528\u539f\u8bed<\/strong>: 2026-06 \u65b0\u8d8b\u52bf\u2014\u2014\u300c\u5b9e\u9a8c \u2192 \u6d4b\u91cf \u2192 \u4fdd\u7559\/\u56de\u6eda\u300d\u5faa\u73af\u5df2\u8131\u79bb ML \u4e13\u7528\uff0c\u6cdb\u5316\u4e3a\u53ef\u590d\u7528\u57fa\u7840\u8bbe\u65bd\uff1api-autoresearch \u53ef\u4f18\u5316\u4efb\u610f\u53ef\u91cf\u5316\u6307\u6807\uff08\u6d4b\u8bd5\u901f\u5ea6\u3001bundle \u4f53\u79ef\u3001Lighthouse \u5206\u6570\uff09\uff0c\u800c\u60f3\u6cd5\u2192\u8bba\u6587\u7cfb\u7edf\uff08AutoResearchClaw\uff09\u4e0e\u957f\u5468\u671f\u534f\u8bae\uff08Deli_AutoResearch\uff09\u628a\u81ea\u4e3b\u65f6\u957f\u4ece\u5206\u949f\u7ea7\u63a8\u5411\u6570\u5468\u7ea7<\/li><li><strong>RL \u7cfb\u7edf\u5c42 + \u5728\u7ebf\u5e38\u9a7b RL<\/strong>: 2026-07 \u65b0\u8d8b\u52bf\u2014\u2014RL \u6548\u7387\u6b63\u72ec\u7acb\u6210\u4e00\u4e2a\u57fa\u7840\u8bbe\u65bd\u5c42\uff1aSnowflake Arctic RL\uff08ZoRRo prompt \u53bb\u91cd\uff0c\u7aef\u5230\u7aef 3.5x \u52a0\u901f\uff09\u4e0e Google Tinker \u517c\u5bb9\u7684 OpenRL\uff08K8s \u539f\u751f\u3001\u591a\u4efb\u52a1 GPU \u6253\u5305\uff09\u628a RL \u7cfb\u7edf\u4f18\u5316\u4e0e\u7b97\u6cd5\u89e3\u8026\uff1b\u800c OpenClaw-RL \u628a<strong>\u5728\u7ebf\u90e8\u7f72\u4e2d\u7684\u771f\u5b9e\u5bf9\u8bdd<\/strong>\u53d8\u6210\u5e38\u9a7b\u8bad\u7ec3\u4fe1\u53f7\u2014\u2014RL \u6b63\u4ece\u79bb\u7ebf\u6279\u5904\u7406\u4f5c\u4e1a\u8f6c\u5411&#8221;\u90e8\u7f72\u5373\u8bad\u7ec3&#8221;\u7684\u5e38\u9a7b\u5f0f\u3001\u6301\u7eed\u5b66\u4e60\u670d\u52a1<\/li><li><strong>\u5728\u7ebf\u7b56\u7565\u84b8\u998f\uff08OPD\uff09\u8d70\u5411\u4e3b\u6d41<\/strong>: 2026-07 \u65b0\u8d8b\u52bf\u2014\u2014OPD\uff08\u5b66\u751f\u751f\u6210\u3001\u6559\u5e08\u5728\u7ebf\u6253\u5206\uff09\u6b63\u6210\u4e3a\u4ecb\u4e8e SFT \u4e0e RL \u4e4b\u95f4\u7684\u6807\u51c6\u540e\u8bad\u7ec3\u9636\u6bb5\uff1bEasyOPD \u5728 verl \u4e4b\u4e0a\u7edf\u4e00 10+ \u79cd OPD \u65b9\u6cd5\uff08\u8de8 tokenizer\u3001\u81ea\u84b8\u998f\u3001\u6b65\u9aa4\u7ea7\uff09\uff0c\u4e00\u884c YAML \u5373\u53ef\u5207\u6362\uff0c\u8ba9\u84b8\u998f\u50cf\u5fae\u8c03\u4e00\u6837\u53ef\u914d\u7f6e<\/li><li><strong>RL \u62d3\u5c55\u5230\u65b0\u8f7d\u4f53 + \u5168\u5b66\u79d1 Autoresearch<\/strong>: RL \u540e\u8bad\u7ec3\u6b63\u7a81\u7834\u81ea\u56de\u5f52\u6587\u672c\u6a21\u578b\u2014\u2014Gen-Verse \u7684 dLLM-RL\/TraceRL \u628a\u8f68\u8ff9\u611f\u77e5 RL \u5e26\u5230<strong>\u6269\u6563\u8bed\u8a00\u6a21\u578b<\/strong>\uff08SOTA TraDo \u7cfb\u5217\uff09\uff0cvLLM \u5b98\u65b9\u7684 vime \u628a RL \u540e\u8bad\u7ec3\u6807\u51c6\u5316\u8fdb\u63a8\u7406\u5f15\u64ce\u751f\u6001\uff1b\u4e0e\u6b64\u540c\u65f6\uff0cautoresearch \u5faa\u73af\u4ece ML \u6cdb\u5316\u4e3a\u5168\u5b66\u79d1\u79d1\u7814\u5de5\u4f5c\u53f0\uff08OpenScience\uff1a\u6587\u732e \u2192 \u5047\u8bbe \u2192 \u5b9e\u9a8c \u2192 \u8bba\u6587\uff0c\u8986\u76d6 ML\/\u751f\u7269\/\u7269\u7406\/\u5316\u5b66\uff09<\/li><li><strong>RL \u6846\u67b6\u5206\u5316\u4e3a\u300c\u6a21\u6001\u4e13\u7528\u77e9\u9635\u300d+ \u4f01\u4e1a\u7ea7 RL<\/strong>: 2026-08 \u65b0\u8d8b\u52bf\u2014\u2014verl \u751f\u6001\u6b63\u5f0f\u88c2\u53d8\u4e3a\u4e09\u6761\u4e13\u7ebf\uff1averl\uff08\u6587\u672c\uff09\u3001<strong>VeRL-Omni<\/strong>\uff08\u6269\u6563\u4e0e\u5168\u6a21\u6001\u751f\u6210\u6a21\u578b\uff09\u3001<strong>verl-vla<\/strong>\uff08\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c\u673a\u5668\u4eba\u7b56\u7565\uff09\uff0cRL \u540e\u8bad\u7ec3\u7531\u6b64\u4ece\u7eaf\u6587\u672c\u6269\u5c55\u5230\u56fe\u50cf\/\u89c6\u9891\/\u97f3\u9891\u751f\u6210\u4e0e\u5b9e\u4f53\u673a\u5668\u4eba\u63a7\u5236\uff1b\u4e0e\u6b64\u540c\u65f6 LMSYS\/SGLang \u56e2\u961f\u7684 <strong>Miles<\/strong>\uff08v0.1\uff0c2026-08-18\uff09\u628a RL \u63a8\u5411\u4e07\u4ebf\u53c2\u6570 MoE \u7684\u751f\u4ea7\u573a\u666f\u2014\u2014\u7cbe\u5ea6\u3001\u7a33\u5b9a\u6027\u3001\u53ef\u89c2\u6d4b\u6027\u4e0e\u541e\u5410\u540c\u7b49\u91cd\u8981<\/li><li><strong>\u81ea\u4e3b\u7814\u7a76\u4ece\u300c\u4ea7\u51fa\u8bba\u6587\u300d\u8f6c\u5411\u300c\u7d2f\u79ef\u77e5\u8bc6\u300d<\/strong>: \u6700\u65b0\u4e00\u6279\u7814\u7a76 Agent \u90fd\u5728\u653b\u540c\u4e00\u4e2a\u5931\u6548\u70b9\u2014\u2014\u6bcf\u6b21\u5b9e\u9a8c\u90fd\u4ece\u96f6\u5f00\u59cb\u3002\u4eba\u5927 NLPIR \u7684 <strong>Arbor<\/strong> \u7528 Hypothesis-Tree Refinement \u628a\u300c\u5047\u8bbe \u2192 \u4ea7\u7269 \u2192 \u8bc1\u636e \u2192 \u6d1e\u89c1\u300d\u6301\u4e45\u5316\u6210\u6811\u5e76\u5411\u4e0a\u4f20\u64ad\u6559\u8bad\uff08\u540c\u7b49\u9884\u7b97\u4e0b\u5e73\u5747\u76f8\u5bf9\u589e\u76ca\u8fbe Codex \/ Claude Code \u7684 2.5 \u500d\uff09\uff1b<strong>Sibyl-AutoResearch<\/strong> \u5219\u76f4\u63a5\u4e3b\u5f20\u300c\u81ea\u4e3b\u7814\u7a76\u9700\u8981\u7684\u662f\u81ea\u8fdb\u5316\u8bd5\u9519 Harness\uff0c\u800c\u4e0d\u662f\u8bba\u6587\u751f\u6210\u5668\u300d\u2014\u2014\u53cd\u590d\u51fa\u73b0\u7684\u6d41\u7a0b\u6027\u5931\u8d25\u4f1a\u6539\u5199 Harness \u81ea\u8eab<\/li><li><strong>Agent Harness \u672c\u8eab\u5f00\u6e90\u5316\u3001\u53ef\u63d2\u62d4\u5316<\/strong>: DeepSeek \u5f00\u6e90\u81ea\u5bb6 Harness\uff08\u4e24\u5929 10 \u4e07 star\uff09\uff0c\u6a21\u578b\u9002\u914d\u5668\u3001\u5de5\u5177\u6ce8\u518c\u8868\u3001\u6c99\u7bb1\u4e43\u81f3 <strong>Agent \u4e3b\u5faa\u73af\u672c\u8eab<\/strong>\u90fd\u662f\u63d2\u4ef6\uff1b\u7ee7 Claw Code \u7684 Clean-Room \u91cd\u5199\u4e4b\u540e\uff0cHarness \u5c42\u6b63\u4ece\u5382\u5546\u9501\u5b9a\u53d8\u4e3a\u901a\u7528\u5f00\u6e90\u57fa\u7840\u8bbe\u65bd\uff1b\u800c Terminal-Bench 2.1\uff08\u65af\u5766\u798f + Laude\uff0c89 \u4e2a\u4efb\u52a1\u542b\u6a21\u578b\u8bad\u7ec3\uff09\u6b63\u6210\u4e3a\u8de8 Harness \u7684\u5171\u540c\u6807\u5c3a<\/li><\/ol>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"\u76f8\u5173-awesome-\u5217\u8868\">\u76f8\u5173 Awesome \u5217\u8868<\/h2>\n\n\n\n<ul><li><a href=\"https:\/\/github.com\/wasiahmad\/Awesome-LLM-Synthetic-Data\">Awesome LLM Synthetic Data<\/a> \u2014 \u5408\u6210\u6570\u636e\u7efc\u8ff0<\/li><li><a href=\"https:\/\/github.com\/Tebmer\/Awesome-Knowledge-Distillation-of-LLMs\">Awesome Knowledge Distillation of LLMs<\/a> \u2014 \u77e5\u8bc6\u84b8\u998f\u7efc\u8ff0<\/li><li><a href=\"https:\/\/github.com\/EnnengYang\/Awesome-Model-Merging-Methods-Theories-Applications\">Awesome Model Merging<\/a> \u2014 \u6a21\u578b\u5408\u5e76\u7efc\u8ff0<\/li><li><a href=\"https:\/\/github.com\/pprp\/Awesome-LLM-Quantization\">Awesome LLM Quantization<\/a> \u2014 \u91cf\u5316\u7efc\u8ff0<\/li><li><a href=\"https:\/\/github.com\/sihyeong\/Awesome-LLM-Inference-Engine\">Awesome LLM Inference Engine<\/a> \u2014 \u63a8\u7406\u5f15\u64ce\u5bf9\u6bd4<\/li><li><a href=\"https:\/\/github.com\/mlabonne\/llm-datasets\">LLM Datasets<\/a> \u2014 \u8bad\u7ec3\u6570\u636e\u96c6\u96c6\u5408<\/li><li><a href=\"https:\/\/github.com\/predibase\/llm_distillation_playbook\">LLM Distillation Playbook<\/a> \u2014 \u84b8\u998f\u5b9e\u8df5\u6307\u5357<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<h2 id=\"\u63a8\u8350\u7ec4\u5408\u76f4\u63a5\u53ef\u7528\">\u63a8\u8350\u7ec4\u5408\uff08\u76f4\u63a5\u53ef\u7528\uff09<\/h2>\n\n\n\n<h3 id=\"-\u6700\u5b8c\u6574\u81ea\u52a8\u5316\u65b9\u6848\">\ud83e\udd47 \u6700\u5b8c\u6574\u81ea\u52a8\u5316\u65b9\u6848<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>HuggingFace Skills + Claude Code + Unsloth + W&amp;B<\/code><\/pre>\n\n\n\n<p>\u7528\u81ea\u7136\u8bed\u8a00\u6307\u6325 Claude Code\uff0c\u901a\u8fc7 HF Skills \u8c03\u7528 Unsloth \u8bad\u7ec3\uff0cW&amp;B \u8ffd\u8e2a\u5b9e\u9a8c\u3002<\/p>\n\n\n\n<h3 id=\"-\u6700\u8f7b\u91cf\u81ea\u4e3b\u5b9e\u9a8c\u65b9\u6848\">\ud83e\udd48 \u6700\u8f7b\u91cf\u81ea\u4e3b\u5b9e\u9a8c\u65b9\u6848<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>AutoResearch + nanochat\uff08\u5355 GPU \u5373\u53ef\u542f\u52a8\uff09<\/code><\/pre>\n\n\n\n<p>\u7761\u524d\u542f\u52a8\uff0c\u9192\u6765\u6536\u83b7 100 \u6b21\u81ea\u4e3b\u5b9e\u9a8c\u7ed3\u679c\u3002<\/p>\n\n\n\n<h3 id=\"-\u6700\u7075\u6d3b\u751f\u4ea7\u65b9\u6848\">\ud83e\udd49 \u6700\u7075\u6d3b\u751f\u4ea7\u65b9\u6848<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>Axolotl \/ LlamaFactory + OpenRLHF + Optuna + MLflow<\/code><\/pre>\n\n\n\n<p>YAML \u914d\u7f6e\u8bad\u7ec3 + \u81ea\u52a8\u8d85\u53c2\u6570\u641c\u7d22 + \u5b9e\u9a8c\u8ffd\u8e2a\u3002<\/p>\n\n\n\n<h3 id=\"-2026-sota-rl-\u8bad\u7ec3\u65b9\u6848\">\ud83c\udfc5 2026 SOTA RL \u8bad\u7ec3\u65b9\u6848<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>verl \/ OpenRLHF + vLLM \/ SGLang + Reasoning Gym + W&amp;B<\/code><\/pre>\n\n\n\n<p>\u6700\u5148\u8fdb\u7684 RL \u8bad\u7ec3 + \u5feb\u901f\u63a8\u7406\u5f15\u64ce + \u4e30\u5bcc\u73af\u5883 + \u5b9e\u9a8c\u8ffd\u8e2a\u3002<\/p>\n\n\n\n<h3 id=\"-\u5408\u6210\u6570\u636e\u5168\u6d41\u7a0b\u65b9\u6848\">\ud83c\udfc5 \u5408\u6210\u6570\u636e\u5168\u6d41\u7a0b\u65b9\u6848<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>Distilabel \/ Magpie \u2192 NeMo Curator \u2192 Unsloth \/ TRL \u2192 DeepEval \/ LMMs-Eval<\/code><\/pre>\n\n\n\n<p>\u89c4\u6a21\u5316\u751f\u6210\u6570\u636e \u2192 GPU \u52a0\u901f\u6e05\u6d17 \u2192 \u9ad8\u6548\u8bad\u7ec3 \u2192 \u5168\u9762\u8bc4\u6d4b\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator\"\/>\n\n\n\n<p><em>\u672c\u62a5\u544a\u57fa\u4e8e 2026 \u5e74 9 \u6708\u516c\u5f00\u8d44\u6e90\u6574\u7406\uff082026-09-07 \u66f4\u65b0\uff09\uff0c\u9879\u76ee\u72b6\u6001\u53ef\u80fd\u968f\u65f6\u95f4\u53d8\u5316\u3002\u5efa\u8bae\u5b9a\u671f\u68c0\u67e5\u5404\u9879\u76ee GitHub \u83b7\u53d6\u6700\u65b0\u72b6\u6001\u3002<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6458\u81ea\uff1aGithub https:\/\/github.com\/BinHPdev\/awesome-algorithm &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2026\/09\/08\/technical-report-on-automated-resource-management-for-large-model-training\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">\u5927\u6a21\u578b\u81ea\u52a8\u5316\u8bad\u7ec3\u8d44\u6e90\u6280\u672f\u62a5\u544a<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[4],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31650"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=31650"}],"version-history":[{"count":2,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31650\/revisions"}],"predecessor-version":[{"id":31652,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31650\/revisions\/31652"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=31650"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=31650"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=31650"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}