aifollow.news
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time AI score54

Google and Penn State study uses edit logs to help AI agents adapt to a new modelMachine translation

AI introduction

A Google and Penn State paper proposes Self-Improvement Memory (SIM), which logs the problem, reasoning and measured result behind each AI-made change to the code that runs an agent’s prompts and tools. After a model switch, the optimizer uses the log to decide which edits to keep or revise. In the reported workspace tasks, the switched setup scored 84.75 with SIM, versus 73.42 without the log.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

谷歌和宾夕法尼亚州立大学的一篇新论文发现,当一个 AI 改写为另一个 AI 提供指令和工具的代码时,记录每次修改的原因有助于让改进效果延续到新模型上。

大语言模型如今可以改写智能体的运行框架,即运行提示词和工具的代码。更换模型或基准测试后,改写后的框架得分可能低于原版,而其代码又无法说明当初为何做出这些修改。

自我改进记忆(Self-Improvement Memory,简称 SIM)会记录每次修改所针对的问题、修改理由及测得的结果。优化器在新配置下读取这份日志,以决定保留或修订哪些旧修改。

将一个 Claude Opus 5.5 运行框架迁移到 Gemini 3.8 Flash 的工作区任务时,SIM 使其得分从 68.41 提高到 84.75,而没有日志时为 73.42。在 10 次重复的编码运行中,SIM 有 9 次得分领先,平均领先 2.36 分。

记录每次运行框架修改的原因及其得分,并将这份日志带到新模型上。

Source: Rohan Paul · x.com

Research
Found an error? Send a correction