Reproduce “ReDel: A Toolkit for LLM-Powered Recursive Multi-Agent Systems” 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

Andrew Zhu, Liam Dugan, Chris Callison-Burch

An open-source toolkit named ReDel was developed to enable the construction and analysis of recursive multi-agent systems powered by Large Language Models. The toolkit's recursive systems achieved higher performance than single-agent baselines and some state-of-the-art models on complex benchmarks, while also identifying prevalent failure modes like 'overcommitment' and 'undercommitment'.

Key insights

  • Recursive delegation, where LLMs dynamically decompose and delegate tasks, consistently leads to improved performance in multi-agent systems compared to single-agent baselines across diverse benchmarks.
  • The performance benefits of recursive delegation scale with the underlying LLM's capabilities, suggesting greater improvements with more advanced models.
  • Common failure modes in recursive multi-agent systems, namely 'overcommitment' (agents tackling overly complex tasks themselves) and 'undercommitment' (agents repeatedly delegating without progress), were systematically identified and quantified.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.