[论文] Does Multi-Agent Collaboration Reduce Hallucination in Clinical Tasks?

论文概要

研究领域: ML 作者: Won Seok Jang, Harry Bhaskar Joshi, Hong Yu 发布时间: 2026-10-22 arXiv: 2610.27027

中文摘要

多智能体LLM系统在临床任务上表现更好，但这种提升常被误读为系统更可信的证据。然而，这是两个不同的主张。临床医生依据所述推理而不仅仅是结论来采取行动，因此协作是否减少幻觉、以及什么因素决定其效果，必须在推理过程中进行衡量。然而大多数评估只对最终答案打分。我们评估了四个系统（单LLM基线和三种多智能体拓扑结构），跨越四个模型和三项临床任务，按照事实性和忠实度分类法对每个系统生成的推理依据中的幻觉进行评分，并与任务指标一并报告。结果表明，协作换来的是任务性能而非更低的幻觉，且没有哪种拓扑结构 reliably 优于其他。效果因模型而异：在36个模型-系统-任务对比中，12个差异足够大——四次降低、八次升高，没有模型同时出现在两侧。决定幻觉水平和变化方向的是模型本身，而非拓扑结构或智能体数量。一个自纠正实验与此一致：在看到被标记的错误后，除Llama-3.2-3B-Instruct外的所有模型都减少了幻觉；协作降低幻觉的那个模型获益最多，而Llama-3.2-3B-Instruct（协作放大其幻觉最多的模型）完全没有减少幻觉。我们发现协作重新分配了其模型已有的能力，但不能提供它们不具备的能力。

原文摘要

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...

#论文 #arXiv #ML #小凯