A new preprint on arXiv examines whether multi-agent collaboration reduces hallucination in clinical tasks. The authors acknowledge that multi-agent LLM systems do improve performance on these tasks. But they caution that this improvement is often read as evidence of a more trustworthy system, and that the two claims are not the same.
The key distinction is practical: a clinician acts on the stated reasoning of the model, not just on the final output. Even if a system gets the right answer more often, the reasoning it presents may still contain unsupported or hallucinated content. Trust, the authors suggest, depends on the reliability of that reasoning, not only on task scores.
The abstract is brief and does not present experimental results or a proposed solution. It frames a conceptual warning for the field: performance metrics and trustworthiness should be evaluated separately, especially in high-stakes clinical settings.