Research map / Large language models / Language, safety and interpretability

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

AAAI 2025 · Large language models › Language, safety and interpretability · approach: Data and hallucination detection · 13 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Uses supervised LLM self-reflection (SR^3) with Llama 3 70B to generate positive and negative reasoning for news, then trains a BERT-based news-reasoning consistency model (NRFE) and a reasoning-free distilled student (NRFE-D) for fake news detection.

Problem

Whether LLM knowledge hallucinations can be exploited as negative reasoning samples to improve fake news detection, given that hallucinations are usually seen as harmful.

Approach

Iteratively prompt an LLM with label-aware polarity and confidence-score constraints (self-reinforced reasoning rectification) to produce both correct positive reasoning and deliberately wrong negative reasoning, then learn news-reasoning semantic consistency with dual BERT encoders and distill into a news-only student.

Key results

• NRFE-D reaches 0.857 accuracy and 0.847 macro F1 on PolitiFact, versus 0.846 / 0.834 for the strongest baseline MUSER.
• NRFE-D reaches 0.958 accuracy on Twitter-15 and 0.971 on Twitter-16, the best among compared methods.
• Zero-shot/CoT open LLMs (Llama 3, Gemma 2, Mistral) score far below fine-tuned encoders, e.g. Llama 3 at 0.769 accuracy on PolitiFact.

Measured effects reported in the paper

Surprising findings

• Open-source LLMs prompted with CoT and zero-shot perform worse than fine-tuned BERT-style encoders on all three datasets.

Limitations

• Negative reasoning is generated under label supervision, so reasoning is label-implicated and the teacher cannot be used at test time.
• Hyperparameters such as margin, polarity threshold and iteration count are fixed without reported sensitivity analysis.
• Evaluated on only three English datasets with small-scale baselines and no reported variance.

Experimental setup

Datasets: PolitiFact Twitter-15 Twitter-16

Benchmarks: PolitiFact Twitter-15 Twitter-16

Metrics: accuracy macro F1 precision recall F1

Baselines: Llama 3 Gemma 2 Mistral BERT RoBERTa ALBERT HSA BiGRU MUSER

Builds on: chain-of-thought knowledge distillation reverse KL loss

Closest papers by meaning

• Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models (AAAI 2026 · similarity 0.87)
• INSIDE: LLMs' Internal States Retain the Power of Hallucination Detection (ICLR 2024 · similarity 0.87)
• Attributive Reasoning for Hallucination Diagnosis of Large Language Models (AAAI 2025 · similarity 0.87)
• Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.87)
• LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.87)
• Zero-resource Hallucination Detection for Text Generation via Graph-based Contextual Knowledge Triples Modeling (AAAI 2025 · similarity 0.86)