Survey on Hallucination in Reasoning Large Language Model: Evaluation, Taxonomy, Intervention, and Open Issues

In recent years, reasoning large language models (LLMs) have seen increasingly widespread adoption in the field of education, particularly demonstrating substantial potential in tasks involving complex text comprehension. However, these LLMs are susceptible to a critical yet often overlooked issue: hallucinations within the reasoning process—instances where the model outputs a correct final answer while its underlying reasoning chain contains fabricated, inconsistent, or logically flawed content. Such hallucination phenomena in Chain-of-Thought (CoT) processes pose serious challenges to the reliability of educational applications. To address this issue, this study proposes a systematic research framework comprising dataset construction, multi-model CoT evaluation, and hallucination classification and quantification. Utilizing the whole-book reading dataset aligned with the junior secondary Chinese language curriculum, we conduct a comparative evaluation of six leading domestic and international LLMs, including ChatGPT o1 and DeepSeek-R1. Key findings include:(1) Hallucinations in CoT are prevalent across all tested models, with ChatGPT-o1 exhibiting a distinctive high accuracy–high hallucination pattern;(2) Hallucinations are both task- and genre-dependent: narrative texts, particularly novels, tend to trigger higher hallucination indices due to long-range dependencies and implicit cultural references. Tasks involving logical reasoning, linguistic feature analysis, and detail extraction show the highest hallucination rates, revealing model weaknesses in handling long-tail knowledge;(3) Hallucinations typically follow a progressive generative pattern: Information misreading → Comprehension deviation → Content fabrication → Logical instability. To mitigate these issues, we propose two targeted intervention strategies: uncertainty-based abstention and model-to-model correction. These approaches offer practical pathways toward enhancing the trustworthiness and educational applicability of reasoning LLMs.