To address the proliferation of fake news on digital platforms, this paper proposes Debate-to-Detect (D2D), a novel fake news detection framework that overcomes the limitations of existing static classification methods. D2D reframes fake news detection as a structured adversarial debate using a multi-agent debate (MAD) approach. Each agent is assigned a specific domain profile and undergoes a five-stage debate process: open speech, rebuttal, free discussion, closed speech, and judgment. Beyond simple binary classification, the paper introduces a multidimensional evaluation mechanism that evaluates each argument across five dimensions: factuality, source credibility, inference quality, clarity, and ethical considerations. Experimental results using GPT-4o demonstrate significant performance improvements over existing methods. A case study demonstrates D2D's ability to iteratively refine evidence and enhance decision-making transparency.
