Multi-Agent Topologies Raise Clinical Task Scores While Hallucination Rates in Rationales Track Model Identity

The arXiv study evaluated a single-LLM baseline against three multi-agent topologies on four models and three clinical tasks. Hallucination was scored inside generated rationales under a joint factuality-faithfulness taxonomy rather than final-answer accuracy alone. Task metrics rose with collaboration, yet rationale hallucination showed no consistent drop. Model fixed effects dominated: Llama-3.2-3B-Instruct amplified errors most, while the single model that benefited from agents also gained most from explicit self-correction on its own flagged mistakes.

Prior clinical LLM evaluations, such as those on Med-PaLM 2 and Med-Gemini, report aggregate accuracy without dissecting rationale faithfulness. This paper isolates the gap: performance gains redistribute existing model capability rather than add missing verification steps. Self-correction experiments confirm the pattern; every model except Llama-3.2-3B-Instruct lowered hallucination when shown its own errors, indicating collaboration merely surfaces what the base model can already correct.

Operationally, hospitals cannot assume multi-agent scaffolding will lower hallucination risk in deployed clinical assistants. Model selection and post-generation verification remain the binding constraints. Future work must report per-rationale hallucination alongside task scores or risk overestimating safety margins.

Deployment timelines that assume agent ensembles will reach regulatory-grade factuality within 12 months lack supporting evidence from the reported distributions.