Latent Debate: surrogate framework for Interpreting LLM Thinking

This paper introduces Latent Debate, a novel framework designed to interpret the internal "thinking" processes and address hallucinations in Large Language Models (LLMs). Unlike external methods that rely on multiple models debating, Latent Debate uses implicit internal arguments—supporting and attacking signals—arising within a single model during a single inference. This framework utilizes a Quantitative Bipolar Argumentation Framework (QBAF) as a "thinking module" to aggregate these internal arguments, successfully serving as a transparent and faithful structured surrogate model for LLM True/False predictions. Empirical analysis demonstrates that this debate pattern is strongly predictive of hallucinations, particularly when intense internal conflicts occur in the middle layers of the LLM architecture.