Reproduce “Generalizable Self-Evolving Memory for Automatic Prompt Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Guanbao Liang, Yuanchen Bei, Sheng Zhou, Yuheng Qin, Huan Zhou, Bingxin Jia, Bin Li, Jiajun Bu
Researchers from Zhejiang University, University of Illinois Urbana-Champaign, and Alibaba Group developed MemAPO, a framework that employs a generalizable, self-evolving dual-memory system for automatic prompt optimization. It achieved an average accuracy of 70.7% on GPT-4o-mini and 70.6% on Qwen3-8B across heterogeneous tasks, improving over baselines while reducing optimization costs by over 55%.
Key insights
- Framing prompt optimization as a process of continual knowledge accumulation, rather than a static search, enables enhanced generalization and efficiency across diverse tasks.
- A structured dual-memory system, explicitly separating successful strategies (positive guidance) and recurrent error patterns (negative constraints), provides comprehensive and effective guidance for LLMs.
- Implementing a self-evolving mechanism within the target LLM reduces reliance on external, more capable optimizers, fostering autonomous improvement and reducing operational costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.