Mirages of Logic: A Survey of Chain-of-Thought Reasoning Hallucinations

ACL ARR 2026 January Submission5642 Authors

Keywords: Large Language Models, Hallucination, Chain-of Thought Reasoning

Abstract: Current research on hallucination focuses mainly on factuality and faithfulness errors in final outputs. However, rapid advances in Chain-of-Thought (CoT) and Large Reasoning Models (LRMs) have improved problem-solving performance while exposing a key weakness: \textbf{Reasoning Hallucination}, characterized by invalid logical transitions or fabricated steps within the reasoning process that can deceptively occur even when the final answer is correct. This paper reviews this emerging challenge, shifting hallucination analysis from output correctness to reasoning processes. We first note a core difficulty that causes reasoning hallucination: long, fluent chains often conceal internal errors, complicating reliability assessment. We then formalize reasoning hallucination and propose a taxonomy of four types: Premise, Operation, Logic, and Conclusion Hallucination. We also examine failure sources and survey mitigation strategies for training and inference. Overall, this work charts a path from answer-level evaluation to transparent, process-reliable reasoning systems.

Paper Type: Long

Research Area: Language Models

Research Area Keywords: chain-of-thought, hallucination, reasoning reliability

Contribution Types: Model analysis & interpretability, Surveys

Languages Studied: English

Submission Number: 5642