Reproduce “Cross-Task Experiential Learning on LLM-based Multi-Agent Collaboration” 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
Yilong Li, Chen Qian, Yu Xia, Ruijie Shi, Yufan Dang, Zihao Xie, Ziming You, Weize Chen and 7 more
The Multi-Agent Cross-Task Experiential Learning (MAEL) framework improves LLM-based multi-agent system performance by enabling agents to accumulate and retrieve high-reward experiences from past tasks. This approach leads to more efficient and higher-quality solutions, demonstrating consistent improvements in solution quality, token consumption, and convergence iterations across various benchmarks.
Key insights
- Quantifying the quality of each reasoning step with rewards and storing these as experiences enables agents to learn from and transfer knowledge across tasks, enhancing efficiency and solution quality.
- Integrating experience retrieval allows MAS to move beyond myopic, scratch-based problem-solving towards an adaptive, experience-driven paradigm.
- Localized optimization at individual decision steps through step-wise experience retrieval contributes most to overall performance, suggesting that diverse, context-specific references promote modular and effective problem-solving.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.