精选理由
对于研究智能体自主进化的开发者，这个框架展示了如何让AI通过自我修改实现性能跃升，值得直接尝试复现。
一个名为“self-improving agents”的开源框架允许智能体修改自身的三个核心部分：执行环境（harness）、底层模型权重和记忆层。这些智能体通过自我评估并根据表现进行适应，实现了自我改进。在MLE-Bench基准测试中，该框架超越了MLEvolve、AIRA-dojo等其他自我改进方案，甚至击败了Karpathy的autoresearcher。这标志着智能体自主进化能力的重要突破，为AI系统的持续优化提供了新思路。
原文 · Ate-a-Pi
This is an open-source framework for self-improvin…
This is an open-source framework for self-improving agents.
These agents can modify 3 things:
• Their harness • Their underlying model's weights • Their memory layer
Basically, these agents evaluate themselves and adapt based on how they are doing.
On MLE-Bench, they beat MLEvolve, AIRA-dojo, and other self-improving solutions.
They also beat Karpathy's autoresearcher.