Detecting LLM Hallucinations Through Topological Signatures of Context Sharing

Summary

This study investigates whether the topology of information flow in attention graphs can distinguish hallucinated responses from non-hallucinated ones. It uses Forman-Ricci curvature to identify structural patterns associated with information bottlenecks, then combines semi-local and global information-flow characteristics from attention heads into a single-pass detection method. The approach was evaluated across several large language models and established hallucination-detection benchmarks. The authors report consistent improvements over existing attention-based and multi-response baselines, while maintaining competitive performance across different model architectures. Their analysis indicates that impaired context sharing during causal generation is strongly associated with hallucination. In particular, hallucinated responses show recurring patterns of excessive self-attention, diffuse retrieval of information from earlier tokens, or information over-squashing. These effects are especially pronounced in the final Transformer layer, suggesting that attention-flow structure can provide signals for detecting unreliable generations without requiring multiple responses.