Google 与 Penn State 论文提出 SIM 方法,让 AI 自改代码的经验跨模型迁移
Google 和 Penn State 把 AI 改 agent 代码时的每次修改原因和得分记成日志,换新模型时直接复用。Claude Opus 5.5 的 harness 迁到 Gemini 3.8 Flash 从 68.41 提到 84.75,比不带日志高 11 分多。经常换底座模型的 agent 开发者可以照这个思路改自己的工作流。
Google 与 Penn State 的论文提出 Self-Improvement Memory(SIM)方法。当前 LLM 可以重写 agent 的 harness(运行提示词和工具的代码),但更换模型或基准后得分可能低于原版,且代码中无法体现每次修改的原因。SIM 的做法是为每次编辑记录三个字段:针对的问题、修改理由和实测结果,优化器在新环境下读取这份日志来决定保留或修改哪些旧编辑。
具体数据上,在将 Claude Opus 5.5 的 harness 迁移到 Gemini 3.8 Flash 的工作区任务时,SIM 把得分从 68.41 提升到 84.75,而没有日志的对照组为 73.42。在 10 次重复编码实验中,SIM 有 9 次领先,平均优势 2.36 分。
这项工作的核心洞察是把'自改进'变成可积累的工程资产:编辑日志记录的是因果与结果,而非代码本身。对需要频繁更换底层模型(如从 Claude 迁到 Gemini)的 agent 开发团队,这提供了一条降低迁移成本的具体路径。
New Google and Penn State paper finds that when one AI rewrites the code that gives another AI its instructions and tools, keeping notes on why each change was made helps the gains carry over to a new model.
LLMs can now rewrite an agent's harness, the code that runs prompts and tools. After a model or benchmark switch, it can score below the original, and its code hides why edits were made.
Self-Improvement Memory, or SIM, logs each edit with the problem it targeted, its reasoning and its measured result. The optimizer reads this log on the new setup to decide which old edits to keep or revise.
When a Claude Opus 5.5 harness moved to Gemini 3.8 Flash workspace tasks, SIM lifted it from 68.41 to 84.75, versus 73.42 without the log. Across 10 repeated coding runs, SIM finished ahead 9 times, by 2.36 points on average.
Store why each harness edit was made and how it scored, and carry that log to new models.
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include SIM, Self-Improvement Memory, Claude, Gemini.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).