Connecting the Dots and Playing Devil's Advocate: Graph-Enhanced Evidence Retrieval and Judicial Multi-Agent Debate for Fake News Detection

Authors:

Abstract
The rapid proliferation of fake news severely erodes public trust and social stability, creating an imperative for robust, evidence-based detection systems. While Retrieval-Augmented Generation (RAG) has advanced this field, it remains constrained by the fragmentation and noise inherent in retrieved evidence, as well as the susceptibility of Large Language Models (LLMs) to reasoning hallucinations. To address these issues, we propose Graph-enhanced Evidence Retrieval and Judicial Multi-Agent DEbate (GRADE), a novel framework unifying heterogeneous graph-enhanced retrieval and multi-agent debate. Specifically, to address evidence fragmentation and noise, we leverage meta-paths on a heterogeneous graph to construct structured context, thereby guiding the retrieval of precise and semantically relevant evidence. Complementing this retrieval enhancement, we introduce a judicial multi-agent debate mechanism inspired by courtroom proceedings. In this module, proponent and opponent agents cross-examine the validity of the evidence, while a judge agent delivers the final verdict, effectively mitigating confirmation bias and reasoning errors. Extensive experiments on the LIAR and CoAID datasets demonstrate that GRADE significantly outperforms state-of-the-art baselines.