Multi-agent scientific discovery with large language models faces a fundamental challenge: the assignments agents receive determine both which experiments are pursued and the experiences from which they adapt. This creates a coupled learning problem, where research direction and agent learning are deeply intertwined.
The authors introduce MELD, a framework that uses hierarchical coordination to manage this coupling. By structuring how research assignments are made and how agents adapt, MELD aims to support recursive self-improvement—allowing the system to refine its own research process over time.
The work is announced on arXiv, though the abstract is cut off before method details are given. Still, the core idea highlights a key obstacle in using LLMs for autonomous science: aligning exploration with learning in a way that lets the system improve itself.