Reproduce “Communicating Plans, Not Percepts: Scalable Multi-Agent Coordination with Embodied World Models” 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
Brennen A. Hill, Mant Koh En Wei, Thangavel Jishnuanandh
A study from the University of Wisconsin-Madison and National University of Singapore reveals that multi-agent coordination is substantially improved by communicating future plans via learned world models rather than through purely emergent messages. This "Intention Communication" approach achieved a 96.5% success rate in complex 15x15 partially observable grid environments, drastically outperforming emergent communication which reached only 12.2%.
Key insights
- Explicitly communicating structured future plans, derived from an agent's internal world model, provides a more robust and scalable mechanism for multi-agent coordination compared to unstructured emergent messages.
- Embedding engineered inductive biases, such as those from a predictive world model, helps decouple the learning problem and facilitates more effective credit assignment for communication in MARL.
- Purely emergent communication protocols, especially with limited message capacity, struggle to establish stable semantics and scale effectively to increasing environmental complexity and partial observability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.