SIA：将智能体视为可编辑系统，实现自改进

精选理由

这篇SIA讲一个让智能体自己改进的方法，不只调提示，还能改工具和权重，在LawBench上做到了70.1%，代码也开源了。

AI 摘要

SIA论文提出将智能体视为可编辑系统，更新目标包括工具、解析器、验证器和权重。在LawBench基准上达到70.1%准确率。其CUDA内核运行仅1,017微秒，去噪任务mse_norm为0.289。论文提供公开仓库，支持选择聚焦于工具链或权重。核心结论是自改进智能体的性能取决于验证器质量。

原文 · AlphaSignal

SIA is interesting because it treats agents as edi…

SIA is interesting because it treats agents as editable systems.

Not just prompts.

Tools, parsers, verifiers, harness code, and weights all become update targets.

The paper reports:

> 70.1% on LawBench > - 1,017 µs CUDA kernel > 0.289 mse_norm on denoising

Public repo makes you pick "--focus harness" or "--focus weights".

The paper’s automatic switch is the missing piece. Real lesson: a self-improving agent is only as good as its verifier.