MELD: Recursive Self-Improvement via Hierarchical Coordination for Multi-Agent Scientific Discovery

Key points

• Multi-agent scientific discovery with large language models raises a coupled learning problem: research assignments determine both which experiments are pursued and the experiences from which agents adapt.1
• The Coordinator generates a joint portfolio of search branches and research directives over a shared derivation tree, while Scientists develop hypotheses and execution plans implemented by a shared frozen Executor.1
• Both levels receive alternating GRPO-style parameter updates under a shared smooth best-of-budget objective, allowing team organization and individual research behavior to co-evolve in a recursive self-improvement loop.1
• Our analysis establishes objective alignment for the hierarchy, characterizes conditions for coordination gains over independent parallel discovery and the first-order benefit of private adaptation, and derives a finite-budget regret bound for adapted models.1

Sources (1)

• [1]MELD: Recursive Self-Improvement via Hierarchical Coordination for Multi-Agent Scientific Discovery
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 04:43 AM

  We introduce MELD, a unified hierarchical test-time training framework that jointly adapts a team-level Coordinator policy and agent-level individual Scientist policies within a single discovery process.

  Multi-agent scientific discovery with large language models raises a coupled learning problem: research assignments determine both which experiments are pursued and the experiences from which agents adapt.

Extractive summary: sentences quoted from the sources.