A new hallucination detector doesn't read what an LLM says, it reads how information moves inside it.
Researchers built a method that analyzes the topology of attention graphs, using Forman-Ricci curvature to spot informational bottlenecks between tokens. The single-pass technique captures both local and global patterns in how attention heads route context during generation. Tested across several LLMs and two established hallucination-detection benchmarks, it delivered consistent improvements over existing attention-based and multi-response baseline methods on those benchmarks. Across different model architectures, though, the results were only competitive with existing approaches, not uniformly better.
Most hallucination detectors either sample multiple responses and compare them, or make a weaker single-pass guess, and neither really explains why a model went off the rails. This method points at a mechanism instead: hallucinated responses line up with impaired context sharing between tokens, showing up as over-reliance on self-attention, diffuse retrieval from early tokens, or information over-squashing in the final transformer layer. That is a diagnostic story, not just a flag, which is the more useful contribution here.
Whether that diagnosis leads anywhere near a fix, or just a fancier warning light, is the question this paper leaves open.