Research ThemeMulti Agent Systems and Data IntelligenceAimTo develop a principled framework for representing, communicating, and aggregating uncertainty in large language model (LLM)–based agent systems, thus improving coordination, calibration and safety in multi-agent reasoning.ObjectivesElicit Uncertainty: Develop and compare methods for extracting calibrated uncertainty estimates from individual LLMs.Represent & Communicate: Design message formats that carry uncertainty information between agents (e.g., probabilistic statements, confidence intervals, prediction sets).Propagate & Aggregate: Establish mathematically sound rules for combining and updating uncertainty as messages pass between agents.Evaluate Impact: Create benchmarks and metrics to quantify how uncertainty-aware communication affects accuracy, calibration, and robustness of multi-agent systems.DescriptionState of the art solutions in many complex problem domains are often now “Agentic” systems in which collaborations of specialist LLM-based agents collectively solve problems. For example, in cyber-security, surveillance agents may monitor logs to detect potential exploits, and cooperate with patching agents that fix vulnerabilities, and testing agents that validate the results. In such systems agents communicate through natural-language messages that reflect single “best guesses.” In probabilistic terms, these are point estimates of a model’s belief. As messages are exchanged, uncertainty information is lost, leading to over-confidence, error cascades, and poor coordination under ambiguity.This project investigates how to restore and maintain epistemic and aleatoric uncertainty throughout an agentic pipeline. The first stage will focus on discovering the most reliable way to elicit and perceive uncertainty estimates with single LLMs. The second stage will formalize uncertainty-aware message formats—for example, by attaching confidence scores, probabilistic distributions, or prediction sets to text outputs. The third stage will develop algorithms for propagating and aggregating uncertainties across agents (eg Bayesian, Dempster-Shafer, conformal fusion). Finally, we will build benchmarks that measure team-level uncertainty calibration. Expected outcomes include (i) theoretical understanding of uncertainty composition in LLM collectives, (ii) practical mechanisms for safer, more transparent AI collaboration, and (iii) open-source tools for uncertainty calibration and propagation in agentic frameworks. Further information https://arxiv.org/abs/2306.10193https://aclanthology.org/2021.acl-long.84/ Closing date: Sat, 31/01/2026 - 12:00 Apply now Principal Supervisor Prof Timothy Hospedales Eligibility Minimum entry qualification - an Honours degree at 2:1 or above (or International equivalent) in a relevant science or engineering discipline, possibly supported by an MSc Degree. Further information on English language requirements for EU/Overseas applicants. Funding Full funding is available for this position.
