TRE Provides Training-Free Hallucination Detection for Diffusion LLMs

TRE is a parameter-free, single-run metric that detects hallucinations in diffusion language models by analyzing entropy signals. It extracts spatial and temporal entropy during the decoding process, eliminating the need for additional training or repeated sampling during inference.