EigenTrack: Spectral Activation Feature Tracking for Hallucination and Out-of-Distribution Detection in LLMs and VLMs

2025-09-19

AI Summary

Large language models (LLMs) and vision-language models (VLMs) frequently suffer from hallucinations and out-of-distribution (OOD) errors, yet existing detection methods lack real-time capability, interpretability, or impose prohibitive computational overhead. Method: We propose a lightweight, real-time detection framework leveraging geometric properties of hidden-layer activation spectra. Our approach constructs a compact global spectral signature by fusing spectral entropy, eigenvalue gaps, and KL divergence of the activation covariance matrix, and models temporal dependencies via a lightweight recurrent neural network—without resampling, model fine-tuning, or external data. Contribution/Results: The method achieves millisecond-level latency while preserving contextual and global signal integrity. It delivers interpretable early warnings and significantly outperforms state-of-the-art black-box, gray-box, and white-box baselines on both hallucination and OOD detection, achieving an optimal trade-off between high accuracy and real-time inference.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models
• Natural Language Processing: (Large) Language Models

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Large language models (LLMs) offer broad utility but remain prone to hallucination and out-of-distribution (OOD) errors. We propose EigenTrack, an interpretable real-time detector that uses the spectral geometry of hidden activations, a compact global signature of model dynamics. By streaming covariance-spectrum statistics such as entropy, eigenvalue gaps, and KL divergence from random baselines into a lightweight recurrent classifier, EigenTrack tracks temporal shifts in representation structure that signal hallucination and OOD drift before surface errors appear. Unlike black- and grey-box methods, it needs only a single forward pass without resampling. Unlike existing white-box detectors, it preserves temporal context, aggregates global signals, and offers interpretable accuracy-latency trade-offs.

Problem

• Detects hallucination and out-of-distribution errors in LLMs
• Uses spectral geometry of hidden activations for real-time monitoring
• Tracks temporal representation shifts before surface errors appear

Innovation

• Uses spectral geometry of hidden activations
• Streams covariance-spectrum statistics into classifier
• Single forward pass without resampling needed