Experience Library Sharing and Transfer Across Multi-Agent Domains

TL;DR: What if agents could "borrow" wisdom from other domains? Let's see if an experience library from one type of multi-agent problem helps another, like negotiation data boosting medical QA or vice versa. The initial experiment would involve cross-domain transfer of reasoning trajectories.

Research Question: Can reasoning trajectories and experience libraries developed in one multi-agent domain (e.g., negotiation) be effectively transferred or adapted to enhance self-improvement in another domain (e.g., biomedical QA)?

Hypothesis: There exist transferable reasoning primitives or patterns in multi-agent LLM interactions that, when shared across domains, enable more rapid adaptation and improved performance in novel or data-scarce scenarios.

Experiment Plan: Construct experience libraries from two or more distinct domains (e.g., negotiation, QA, clinical trial design). Develop methods for aligning or mapping reasoning steps/representations between domains (e.g., via embedding alignment or meta-learning). Train or fine-tune multi-agent LLM systems in a target domain with and without imported experience library data from a source domain. Measure performance improvements, sample efficiency, and analyze which types of reasoning transfer most successfully.

References:

• Zhao, W., Yuksekgonul, M., Wu, S., & Zou, J. (2025). SiriuS: Self-improving Multi-agent Systems via Bootstrapped Reasoning. arXiv.org.
• Yue, L., Xing, S., Chen, J., & Fu, T. (2024). ClinicalAgent: Clinical Trial Multi-Agent System with Large Language Model-based Reasoning. ACM International Conference on Bioinformatics, Computational Biology and Biomedicine.