Topological Hallucination Detection via Forman-Ricci Curvature in Attention Graphs

This method identifies hallucinations by analyzing information bottlenecks in attention graphs using Forman-Ricci curvature. By capturing semi-local and global information-flow patterns, the single-pass approach outperforms existing attention-based and multi-response detection baselines.