From AI debate to human-AI consensus: communication protocols for multi-agent collaboration

Collaboration among multiple agents, whether AI models or humans, offers significant potential for improving decision-making and predictive accuracy. In multi-agent settings with private information, communication protocols determine whether and how agents reach consensus, and with what statistical quality. Classical results in Bayesian game theory show that exchanging full posteriors guarantees Bayes-optimal consensus in a single round, but this approach is computationally expensive and impractical for large-scale systems such as large language models (LLMs). Current practice, such as multi-agent debate among LLMs, often relies on sample-based communication, which is efficient but may lead to suboptimal or approximate consensus.

This project addresses two key challenges: (1) developing a theoretical framework to quantify the optimality gap for sample-based communication and designing hybrid protocols that balance accuracy and cost for AI-only collaboration; and (2) extending these protocols to human-AI collaborative environments, such as medical analysis where interpretability, trust, and cognitive efficiency become critical. By integrating insights from Bayesian consensus theory, game-theoretic agreement, theory of mind and modern multi-agent AI systems, this research aims to deliver principled, cost-aware communication mechanisms for a trustworthy collaboration.

You will need to meet the minimum entry requirements for our PhD programme.

Open to candidates who pay UK/home rate fees. See UKCISA for further information.