Paper Overview
Field: ML Authors: Won Seok Jang, Harry Bhaskar Joshi, Hong Yu Published: 2026-10-22 arXiv: 2610.27027
Summary
Multi-agent LLM systems improve performance on clinical tasks, but this improvement is often misread as evidence that the system is more trustworthy. These are two different claims. Clinicians act on stated reasoning, not just conclusions, so whether collaboration reduces hallucination—and what determines when it does—must be measured within the reasoning process itself. Yet most evaluations only score the final answer.
The authors evaluate four systems (a single-LLM baseline and three multi-agent topologies) across four models and three clinical tasks, scoring hallucination inside each system's generated rationale under a factuality and faithfulness taxonomy, and reporting it alongside task metrics.
Key Findings
- Collaboration buys task performance but not lower hallucination as a rule, and no topology reliably outperforms the others.
- Effects vary by model: among 36 model-system-task comparisons, 12 show differences large enough to matter—four reductions and eight increases, with no model appearing on both sides.
- The model itself, rather than topology or agent count, determines hallucination levels and the direction of change.
- A self-correction experiment is consistent with this: after seeing flagged errors, all models except Llama-3.2-3B-Instruct reduced hallucination. The model whose collaboration reduced hallucination benefited the most, while Llama-3.2-3B-Instruct (the model whose hallucination was most amplified by collaboration) showed no reduction at all.
Conclusion
Collaboration redistributes capabilities a model already has, but cannot provide capabilities it lacks.
Original Abstract
Multi-agent LLM systems improve performance on clinical tasks, and that improvement is often read as evidence of a more trustworthy system. However, the two are not the same claim. A clinician acts on the stated reasoning and not only on the conclusion, so whether collaboration reduces hallucination, and what decides when it does, has to be measured in the reasoning too. Yet most evaluations score only the final answer. We evaluate four systems (a single-LLM baseline and three multi-agent topologies) across four models and three clinical tasks, scoring hallucination, under a factuality and faithfulness taxonomy, inside the rationale each system generates and reporting it alongside the task metric. We report that collaboration buys task performance but not lower hallucination as a rule, and...
---
*Auto-collected on 2026-10-22*