Multi-Agent Debate: Analyzing Consensus in Networks of LLM Agents

Large language models (LLMs) are increasingly being deployed in systems of interconnected agents. By giving agents access to their own past chain of thought or those of other agents, accuracy on many tasks can be increased. Yet in multi-agent systems, which consist of interactions between agents, it remains unclear how the network structure of interactions impacts accuracy, collaboration, and consensus. Using tools from complex systems and physics, we analyze the effects of different network graph structures on consensus in networks of debating agents and compare speed versus accuracy tradeoffs. We also consider other factors beyond the structure of interactions, such as the model type, prompt structure, and sampling temperature. We anticipate that our results will have implications for future work designing multi-agent systems, producing accurate responses in multi-agent systems, and reducing test-time compute resources by efficiently structuring debate.