Multi‑agent large language model (LLM) systems often achieve higher scores on clinical tasks, and this improvement is sometimes taken as evidence of greater trustworthiness. However, clinicians rely on the full reasoning chain, not just the final answer, so hallucination must be measured within the generated rationale. We evaluated four systems—a single‑LLM baseline and three multi‑agent topologies—across four models and three clinical tasks, scoring hallucination using a factuality‑and‑faithfulness taxonomy inside each system's rationale and reporting it alongside the task metric. The results show that collaboration boosts task performance but does not consistently reduce hallucination, and no topology is reliably superior. The determining factor is the underlying model rather than the topology or number of agents. Among $36$ model‑system‑task comparisons, only $12$ showed statistically significant differences—$4$ decreases and $8$ increases—and no model exhibited both effects. A self‑correction experiment confirmed this pattern: when shown its own flagged errors, every model except Llama-3.2-3B-Instruct reduced hallucination; the latter’s hallucination was amplified and never decreased. Thus, collaboration merely redistributes what models can already do; it does not provide capabilities they lack.
Review