Study finds multi-agent AI setups improve clinical task scores but do not reliably reduce hallucinated reasoning
Across four models and three clinical tasks, the underlying model mattered more than how the agents were arranged.
A new arXiv paper by Won Seok Jang, Harry Bhaskar Joshi and Hong Yu asks whether having several language-model agents work together reduces hallucination on clinical tasks. The authors note that most evaluations score only the final answer, even though clinicians act on a system's stated reasoning as well as its conclusion. They compared four systems, a single-model baseline and three multi-agent topologies, across four models and three clinical tasks. They scored hallucination inside each generated rationale, using a factuality and faithfulness taxonomy, alongside the usual task metric. They found that collaboration improved task performance but did not, as a rule, reduce hallucination, and no topology was reliably better than another. Of 36 model, system and task comparisons, 12 showed effects large enough to resolve: four decreases in hallucination and eight increases. No model appeared on both sides. The authors conclude that the model, not the collaboration structure, sets the level of hallucination and the direction of any change. The finding matters for hospitals and developers who might treat better benchmark scores from multi-agent systems as evidence of more trustworthy clinical reasoning.