Are Reasoning Models More Prone to Hallucination?

Recently evolved large reasoning models (LRMs) show powerful performance insolving complex tasks with long chain-of-thought (CoT) reasoning capability. Asthese LRMs are mostly developed by post-training on formal reasoning tasks,whether they generalize the reasoning capability to help reduce hallucinationin fact-seeking tasks remains unclear and debated. For instance, DeepSeek-R1reports increased performance on SimpleQA, a fact-seeking benchmark, whileOpenAI-o3 observes even severer hallucination. This discrepancy naturallyraises the following research question: Are reasoning models more prone tohallucination? This paper addresses the question from three perspectives. (1)We first conduct a holistic evaluation for the hallucination in LRMs. Ouranalysis reveals that LRMs undergo a full post-training pipeline with coldstart supervised fine-tuning (SFT) and verifiable reward RL generally alleviatetheir hallucination. In contrast, both distillation alone and RL trainingwithout cold start fine-tuning introduce more nuanced hallucinations. (2) Toexplore why different post-training pipelines alters the impact onhallucination in LRMs, we conduct behavior analysis. We characterize twocritical cognitive behaviors that directly affect the factuality of a LRM: FlawRepetition, where the surface-level reasoning attempts repeatedly follow thesame underlying flawed logic, and Think-Answer Mismatch, where the final answerfails to faithfully match the previous CoT process. (3) Further, we investigatethe mechanism behind the hallucination of LRMs from the perspective of modeluncertainty. We find that increased hallucination of LRMs is usually associatedwith the misalignment between model uncertainty and factual accuracy. Our workprovides an initial understanding of the hallucination in LRMs.