Is LLMs Hallucination Usable? LLM-based Negative Reasoning for Fake News Detection

2025-03-12

AI Summary

Large language models (LLMs) suffer from knowledge hallucinations that lead to unreliable reasoning, undermining their utility in fact-sensitive tasks like fake news detection. Method: This paper introduces SR³, the first supervised self-reinforced reasoning correction framework that repurposes hallucination-induced erroneous (i.e., negative) reasoning as a discriminative signal. It constructs NRFE—a semantic consistency representation model trained on positive/negative news–reasoning pairs—and distills it into a lightweight student model, NRFE-D. The approach integrates reflective reasoning, multi-stage prompting, semantic consistency modeling, and supervised self-reinforcement learning—bypassing label dependency and its associated biases. Results: Evaluated on three benchmark fake news datasets, SR³ significantly outperforms LLM prompting, fine-tuned small models, and existing state-of-the-art methods, demonstrating that leveraging negative reasoning substantially enhances detection robustness and generalization.

Technology Category

• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Common-Sense Reasoning

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

The questionable responses caused by knowledge hallucination may lead to LLMs' unstable ability in decision-making. However, it has never been investigated whether the LLMs' hallucination is possibly usable to generate negative reasoning for facilitating the detection of fake news. This study proposes a novel supervised self-reinforced reasoning rectification approach - SR$^3$ that yields both common reasonable reasoning and wrong understandings (negative reasoning) for news via LLMs reflection for semantic consistency learning. Upon that, we construct a negative reasoning-based news learning model called - emph{NRFE}, which leverages positive or negative news-reasoning pairs for learning the semantic consistency between them. To avoid the impact of label-implicated reasoning, we deploy a student model - emph{NRFE-D} that only takes news content as input to inspect the performance of our method by distilling the knowledge from emph{NRFE}. The experimental results verified on three popular fake news datasets demonstrate the superiority of our method compared with three kinds of baselines including prompting on LLMs, fine-tuning on pre-trained SLMs, and other representative fake news detection methods.

Problem

• Explores usability of LLMs' hallucination for fake news detection.
• Proposes SR^3 for generating negative reasoning via LLMs reflection.
• Develops NRFE model using positive-negative reasoning pairs for learning.

Innovation

• SR^3: Supervised self-reinforced reasoning rectification approach
• NRFE: Negative reasoning-based news learning model
• NRFE-D: Student model for knowledge distillation