Effectiveness of Leveraging Network Topology by Multi-Agent LLMs

Determine whether large language model-based multi-agent systems are able to leverage their communication network topology effectively when solving reasoning tasks that require coordination, self-organization, and multi-round message-passing.

Background

The paper introduces AgentsNet to evaluate multi-agent LLMs on foundational distributed computing tasks that require structured coordination and communication across a network. Prior work shows improvements from multi-agent setups, but these are often assessed on standard reasoning benchmarks that may not explicitly test the use of topology.

The authors highlight uncertainty about whether current multi-agent LLM systems meaningfully exploit network structure, motivating a benchmark grounded in graph-theoretic and distributed systems problems to probe this capability.