Google researchers introduce SIM for continual agent self-improvement
Google Cloud AI, Google DeepMind and the University of Pennsylvania jointly proposed Self-Improvement Memory (SIM), an updatable record of harness changes, their rationale, expected outcomes and observed results. SIM supports continual recursive self-improvement when tasks or models change without altering model weights; reported tests covered programming, text classification, office tasks and engineering design. In 10 paired programming comparisons, SIM improved 9 runs by an average of 2.36 points, while the largest reported cross-model office-task gap was 11.33 points.
Academic Headlines notes that carrying modification history and evaluation results beyond inherited code can help an optimizer reassess changes after a model or benchmark changes, while the experiments remain bounded by fixed interfaces, optimization rounds and limited repeated runs outside programming.

智东西 作者 | 江宇 编辑 | 李水青
智东西11月19日报道,近日,谷歌云AI、谷歌DeepMind与美国宾夕法尼亚州立大学的研究团队联合发布论文《自改进记忆,实现持续递归式自改进》(Self-Improvement Memory Enables Continual Recursive Self-Improvement),提出了一种名为自我改进记忆(Self-Improvement Memory,SIM)的新方法。

▲《自我改进记忆实现持续递归自我改进》
此前,AI虽然能够自主优化大模型外部的运行框架(Harness),但更换模型或任务后,往往需要重新优化。SIM首次将代码修改背后的原因、预期效果和实际结果组织成可持续更新的记忆,让AI能够继承并调整过去的改进经验,实现跨模型、跨任务的持续递归自我改进(CRSI)。
#Self-Improvement Memory#Continual Recursive Self-Improvement#Google Cloud AI#Google DeepMind#University of Pennsylvania#Gemini 2.5 Flash#Gemini 3.5 Flash#Gemini 3.8 Flash#Gemini 4 Argon#Claude Opus 5.5