https://www.hansking.cn/blog A posts created with Next.js and Tailwind.css en-us admin@hansking.cn (HANS) admin@hansking.cn (HANS) Sun, 16 Feb 2025 10:00:00 GMT https://www.hansking.cn/blog/2026/DeepSeek-R1-de-xun-lian-si-lu-yu-kai-yuan-ying-xiang https://www.hansking.cn/blog/2026/DeepSeek-R1-de-xun-lian-si-lu-yu-kai-yuan-ying-xiang [{"value":"DeepSeek-R1 的训练思路与开源影响","url":"#deepseek-r1-的训练思路与开源影响","depth":1},{"value":"一、R1-Zero：先让模型在奖励里寻找能力","url":"#一r1-zero先让模型在奖励里寻找能力","depth":2},{"value":"二、为什么冷启动数据仍然重要","url":"#二为什么冷启动数据仍然重要","depth":2},{"value":"三、可验证奖励是整条路线的地基","url":"#三可验证奖励是整条路线的地基","depth":2},{"value":"四、GRPO 为什么成为重要的配套方法","url":"#四grpo-为什么成为重要的配套方法","depth":2},{"value":"五、推理时计算为什么重新成为能力变量","url":"#五推理时计算为什么重新成为能力变量","depth":2},{"value":"六、蒸馏让成果进入更小的模型","url":"#六蒸馏让成果进入更小的模型","depth":2},{"value":"七、开源影响不只是放出一个权重文件","url":"#七开源影响不只是放出一个权重文件","depth":2},{"value":"八、如何避免把 R1 学成几个口号","url":"#八如何避免把-r1-学成几个口号","depth":2},{"value":"九、我的总结：R1 改变的是问题的问法","url":"#九我的总结r1-改变的是问题的问法","depth":2}] Sun, 16 Feb 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMDeepSeek推理模型强化学习 https://www.hansking.cn/blog/2026/GRPO-ru-men-qun-ti-xiang-dui-ce-le-you-hua https://www.hansking.cn/blog/2026/GRPO-ru-men-qun-ti-xiang-dui-ce-le-you-hua [{"value":"GRPO 入门：群体相对策略优化","url":"#grpo-入门群体相对策略优化","depth":1},{"value":"一、先回顾 PPO 在做什么","url":"#一先回顾-ppo-在做什么","depth":2},{"value":"二、GRPO 的核心直觉","url":"#二grpo-的核心直觉","depth":2},{"value":"三、目标函数在优化什么","url":"#三目标函数在优化什么","depth":2},{"value":"四、奖励函数决定模型究竟学会什么","url":"#四奖励函数决定模型究竟学会什么","depth":2},{"value":"五、组怎么采样，比想象中更重要","url":"#五组怎么采样比想象中更重要","depth":2},{"value":"六、长度偏差是一个隐蔽的坑","url":"#六长度偏差是一个隐蔽的坑","depth":2},{"value":"七、一个简化训练流程","url":"#七一个简化训练流程","depth":2},{"value":"八、训练中应该盯哪些指标","url":"#八训练中应该盯哪些指标","depth":2},{"value":"九、GRPO 和 SFT、PPO 如何配合","url":"#九grpo-和-sftppo-如何配合","depth":2},{"value":"十、我的总结：比较可以替代一部分“绝对判断”","url":"#十我的总结比较可以替代一部分绝对判断","depth":2}] Sun, 02 Mar 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLM推理模型GRPO强化学习 https://www.hansking.cn/blog/2026/Test-timeCompute-yong-geng-duo-tui-li-huan-geng-gao-zheng-que-l https://www.hansking.cn/blog/2026/Test-timeCompute-yong-geng-duo-tui-li-huan-geng-gao-zheng-que-l [{"value":"Test-time Compute：用更多推理换更高正确率","url":"#test-time-compute用更多推理换更高正确率","depth":1},{"value":"一、训练时计算和推理时计算有什么不同","url":"#一训练时计算和推理时计算有什么不同","depth":2},{"value":"二、最简单的方式：Best-of-N","url":"#二最简单的方式best-of-n","depth":2},{"value":"三、自洽性不是简单投票","url":"#三自洽性不是简单投票","depth":2},{"value":"四、搜索与分支：把答案当成一个小型搜索问题","url":"#四搜索与分支把答案当成一个小型搜索问题","depth":2},{"value":"五、动态预算比固定长思考更合理","url":"#五动态预算比固定长思考更合理","depth":2},{"value":"六、推理时计算的成本怎么算","url":"#六推理时计算的成本怎么算","depth":2},{"value":"七、质量收益必须通过对照实验证明","url":"#七质量收益必须通过对照实验证明","depth":2},{"value":"八、什么时候不值得使用更多计算","url":"#八什么时候不值得使用更多计算","depth":2},{"value":"九、生产环境必须设置护栏","url":"#九生产环境必须设置护栏","depth":2},{"value":"十、我的总结：把计算当成预算，而不是信仰","url":"#十我的总结把计算当成预算而不是信仰","depth":2}] Sun, 02 Feb 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLM推理模型Test-time Compute性能优化 https://www.hansking.cn/blog/2026tui-li-mo-xing-xing-qi-wei-shen-me-ce-shi-shi-ji-suan-kai-shi-bian-zhong-yao https://www.hansking.cn/blog/2026tui-li-mo-xing-xing-qi-wei-shen-me-ce-shi-shi-ji-suan-kai-shi-bian-zhong-yao [{"value":"推理模型兴起：为什么测试时计算开始变重要","url":"#推理模型兴起为什么测试时计算开始变重要","depth":1},{"value":"一、为什么单次生成开始遇到瓶颈","url":"#一为什么单次生成开始遇到瓶颈","depth":2},{"value":"二、测试时计算到底在增加什么","url":"#二测试时计算到底在增加什么","depth":2},{"value":"三、推理模型为什么看起来像是在思考","url":"#三推理模型为什么看起来像是在思考","depth":2},{"value":"四、从 Chain-of-Thought 到可验证推理","url":"#四从-chain-of-thought-到可验证推理","depth":2},{"value":"五、为什么这会改变模型训练","url":"#五为什么这会改变模型训练","depth":2},{"value":"六、推理预算是一种新的服务资源","url":"#六推理预算是一种新的服务资源","depth":2},{"value":"七、质量曲线比单点榜单更重要","url":"#七质量曲线比单点榜单更重要","depth":2},{"value":"八、它对模型产品意味着什么","url":"#八它对模型产品意味着什么","depth":2},{"value":"九、它没有解决什么问题","url":"#九它没有解决什么问题","depth":2},{"value":"十、作为开发者应该怎样学习","url":"#十作为开发者应该怎样学习","depth":2},{"value":"十一、我的总结：模型的能力开始有了“运行时维度”","url":"#十一我的总结模型的能力开始有了运行时维度","depth":2}] Sun, 05 Jan 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLM推理模型Test-time Compute发展史 https://www.hansking.cn/blog/2026tui-li-mo-xing-de-si-wei-lian-ping-ce-yu-jie-guo-jiao-yan https://www.hansking.cn/blog/2026tui-li-mo-xing-de-si-wei-lian-ping-ce-yu-jie-guo-jiao-yan [{"value":"推理模型的思维链评测与结果校验","url":"#推理模型的思维链评测与结果校验","depth":1},{"value":"一、先分清三个概念","url":"#一先分清三个概念","depth":2},{"value":"最终结果监督：答案对不对","url":"#最终结果监督答案对不对","depth":3},{"value":"过程监督：中间步骤是否合理","url":"#过程监督中间步骤是否合理","depth":3},{"value":"可验证中间结果：能否由系统复查","url":"#可验证中间结果能否由系统复查","depth":3},{"value":"二、不要把“展示思维链”当成评测的终点","url":"#二不要把展示思维链当成评测的终点","depth":2},{"value":"三、先设计可评测的任务，而不是先挑模型","url":"#三先设计可评测的任务而不是先挑模型","depth":2},{"value":"四、自动校验器是推理评测的核心资产","url":"#四自动校验器是推理评测的核心资产","depth":2},{"value":"数值与公式","url":"#数值与公式","depth":3},{"value":"代码","url":"#代码","depth":3},{"value":"SQL 与数据操作","url":"#sql-与数据操作","depth":3},{"value":"结构化计划","url":"#结构化计划","depth":3},{"value":"五、如何评测中间步骤","url":"#五如何评测中间步骤","depth":2},{"value":"六、自洽采样有用，但不能迷信多数票","url":"#六自洽采样有用但不能迷信多数票","depth":2},{"value":"七、裁判模型只能当辅助裁判","url":"#七裁判模型只能当辅助裁判","depth":2},{"value":"八、生产环境要校验结果，而不只是评测结果","url":"#八生产环境要校验结果而不只是评测结果","depth":2},{"value":"九、我在评测中最常见的三个误区","url":"#九我在评测中最常见的三个误区","depth":2},{"value":"十、我的总结：让结论接受事实的审判","url":"#十我的总结让结论接受事实的审判","depth":2}] Sun, 16 Mar 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLM推理模型评测思维链