Search papers, labs, and topics across Lattice.
This paper introduces Free-Energy Signatures (Fes), a novel spectral descriptor for detecting hallucinations in large language models (LLMs) by analyzing the thermodynamic potentials of attention-derived graph Laplacians. By leveraging the full structure of the Laplacian spectrum rather than relying on a few eigenvalues, Fes demonstrates improved performance in distinguishing between correct and hallucinated outputs. Empirical results show that a lightweight probe utilizing Fes achieves a significant average AUROC improvement over existing attention-spectral baselines, highlighting its effectiveness as a training-free detection method.
Spectral diagnostics reveal that hallucinations in LLMs can be detected with unprecedented accuracy by analyzing the thermodynamic properties of attention graphs.
Hallucination detection in large language models (LLMs) is deployment-critical, and recent work shows that the spectrum of attention-derived graph Laplacians carries strong signal about reasoning quality. Prior spectral diagnostics, however, summarize the Laplacian spectrum by a handful of eigenvalues or hand-picked scalars, leaving most of its structure unused. We propose Free-Energy Signatures (Fes), a spectral descriptor that treats each layer's attention Laplacian as a Hamiltonian and extracts its thermodynamic potentials partition function, free energy, spectral entropy, heat capacity together with the random-matrix-theory (RMT) spectral form factor. We prove three results: (i)~Lipschitz stability of Fes under attention perturbation; (ii)~an expressiveness result showing that Fes enriches finite spectral summaries and approximates moment-derived spectral functionals under explicit regularity and grid-resolution assumptions; and (iii)~a finite-sample PAC bound on the AUROC of a training-free detector built from Fes. Empirically, across six open-weight LLMs and six benchmarks, a lightweight probe on Fes descriptors achieves the strongest aggregate AUROC among attention-spectral baselines, improving over LapEig by $+6.5$ AUROC points and over GoR-4 by $+2.4$ points on average, while requiring no update to the underlying LLM. In the fully unsupervised setting, an RMT-deviation score achieves mean AUROC $0.71$, providing a label-free but weaker detector. A complementary RMT analysis shows that correct generations exhibit more Wigner-Dyson like spectral statistics, whereas hallucinations exhibit more Poisson-like statistics. The anonymized code and config are provided in the supplementary material.
