Reproduce “A Graph Signal Processing Framework for Hallucination Detection in Large Language Models” on OpenResearch

About this paper

Valentin Noël

A Graph Signal Processing framework models transformer layers as dynamic graphs, treating token embeddings as signals, to analyze internal information flow and detect large language model hallucinations. This approach identifies universal spectral patterns for factual reasoning and distinct signatures for different hallucination types, achieving 88.75% detection accuracy, outperforming perplexity-based and SelfCheckGPT baselines.

Key insights

• Factual reasoning in transformer models consistently exhibits a "universal energy mountain" trajectory, where energy builds in middle layers and significantly dissipates at final output layers, converging to low graph frequencies.
• Different types of hallucinations manifest distinct and identifiable spectral fingerprints: logical errors cause strong spectral instability, semantic errors lead to subtle 'connectivity drift' (or collapse in newer models), and substitution errors show intermediate spectral perturbations.
• The Fiedler value, representing algebraic connectivity, consistently increases during factual reasoning, indicating increasingly coherent and robust internal representations.

Summary by alphaXiv.