PRISM: Refracting Hallucinations in Chain-of-Thought Reasoning via Multi-Dimensional Verification

ACL ARR 2026 January Submission2168 Authors

Keywords: Chain-of-Thought reasoning, hallucination mitigation, question comprehension, multi-dimensional verification, commonsense reasoning

Abstract: In recent years, Chain-of-Thought (CoT) verification has emerged as a critical research direction. However, existing approaches largely focus on the quality of intermediate reasoning or final answer correctness, while hallucinations arising from the initial stage of question understanding remain underexplored. To address this gap, we propose a unified framework—PRISM (\textbf{P}rogressive \textbf{R}easoning with \textbf{I}nstructional and \textbf{S}trategic \textbf{M}ulti-dimensional Verification) that jointly tackles all three aspects. We introduce a Commonsense-Augmented Progressive Instructional Reasoning (CPIR) method, designed to alleviate condition hallucination while utilizing commonsense to capture relevance between conditions and questions. Then we develop Multi-Dimensional Heterogeneous Collaborative Verification (MHCV), which strategically validates reasoning chains from multiple perspectives to enhance intermediate reasoning quality and question comprehension, thereby mitigating different types of hallucinations. In addition, we propose a Discard-Weighted Voting mechanism to overcome the limitations of traditional voting methods in multi-dimensional verification. Experimental results demonstrate that PRISM consistently improves verification accuracy across conditions, logical reasoning, and question comprehension, yielding more reliable reasoning chains and higher final-answer accuracy compared to strong CoT baselines.

Paper Type: Long

Research Area: Mathematical, Symbolic, Neurosymbolic, and Logical Reasoning

Research Area Keywords: chain-of-thought, prompting, LLM agents

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 2168