IMPD-MACD: A Comprehensive Multi-Perspective Cognitive Fusion Approach for LLM Hallucination Detection

ACL ARR 2025 February Submission699 Authors

Abstract: In recent years, researchers have observed that LLMs frequently generate false content ("hallucinations") with highly confident tones when answering questions, a phenomenon that severely undermines model reliability. Therefore, timely and accurate detection of LLM-generated hallucinations is crucial. However, existing methods face multiple challenges in hallucination detection: (1) Single-agent methods lack comprehensive identification of different types of hallucinations, and some methods are difficult to apply to black-box models; (2) Multi-agent methods lack clear division of labor and deep interaction, and combined with inherent biases and overconfidence issues, their detection effectiveness is insufficient in complex scenarios. In response, we propose IMPD-MACD, which enhances agent division of labor and collaboration through multiple perspectives and stances, not only significantly improving detection accuracy but also more comprehensively covering different types of hallucinations, thereby enhancing LLM reliability and practicality in diverse scenarios. Extensive experiments demonstrate that our method significantly outperforms the current SOTA (state-of-the-art) approaches across multiple metrics. The project and its associated dataset will be publicly released.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: fact checking, rumor/misinformation detection

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 699