CRAgents for Disease Modelling via Multi-Agent Critique and Refinement

Kelly Yu, Zuoou Li, Paul M. Matthews, Wenjia Bai, Bernhard Kainz, Mengyun Qiao

Keywords: multi-agent systems, automated machine learning, LLM agents

Registration Requirement: Yes

Abstract: Automating disease modelling requires systems that can formulate clinically valid prediction tasks and optimise robust models under low-prevalence settings. However, current multi-agent workflows do not unify these stages, leading to clinically weak target selection and reliance on manual model configuration. We propose CRAgents, a multi-agent framework that couples task formulation with model optimisation within a unified workflow. Built from eight agents, it introduces two reasoning loops: (i) a debate-driven loop (Inner Debate Loop) that performs critique through systematic, clinically grounded evaluation of modelling targets; and (ii) a reflection-driven loop (Outer Reflection Loop) that enables refinement through iterative model improvement based on validation behaviour. We evaluate CRAgents on the UK Biobank dataset, where it outperforms recent agent-based baselines, achieving an AUC of 0.781, demonstrating that the combination of critique and refinement improves reliability and clinical alignment in automated disease modelling.

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 42