成本不到150元，训练出一个媲美DeepSeek-R1和OpenAI o1的推理模型？！ 这不是洋葱新闻，而是AI教母李飞飞、斯坦福大学、华盛顿大学、艾伦人工智能 ...

这种方法可以引导模型进行自我检查，并修正推理过程中的错误，从而提高推理性能。 具体来说，他们构建了一个叫做「s1K」的数据集，由1000个精心筛选的问题组成，每个问题都配有推理轨迹（reasoning traces）和从Gemini Thinking Experimental蒸馏而来的答案。 接着 ...

在 s1 的新工作中，研究人员寻求最简单的方法来实现测试时间扩展。它们构建了一个小型数据集 s1K，其中包含 1000 个问题，并根据三个标准（难度、多样性和质量）与推理轨迹进行配对。 在此基础上，研究人员开发了「预算强制」来控制测试时间计算 ...

Their method centers on two key innovations: the carefully curated s1K dataset comprising 1,000 questions with reasoning traces, selected based on difficulty, diversity, and quality criteria, and a ...

OpenAI’s big rebranding effort brings a new logo and a new typeface, OpenAI sans. OpenAI’s big rebranding effort brings a new logo and a new typeface, OpenAI sans. Emma Roth is a news writer ...

On Monday, three logos were revealed to the public, one for each potential nickname, but NHL fans weren't particularly impressed by the latest round of options. Much of the criticism was focused ...