ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMs

TL;DR

This paper introduces the ICR Probe, a novel method that tracks hidden state dynamics across layers in large language models to reliably detect hallucinations, outperforming existing static approaches.

Contribution

The paper proposes the ICR Score and ICR Probe, focusing on hidden state evolution over layers, providing a more effective and interpretable hallucination detection method for LLMs.

Findings

ICR Score effectively distinguishes hallucinations from non-hallucinations.

ICR Probe achieves superior detection performance with fewer parameters.

Ablation studies reveal insights into hidden state dynamics and model interpretability.

Abstract

Large language models (LLMs) excel at various natural language processing tasks, but their tendency to generate hallucinations undermines their reliability. Existing hallucination detection methods leveraging hidden states predominantly focus on static and isolated representations, overlooking their dynamic evolution across layers, which limits efficacy. To address this limitation, we shift the focus to the hidden state update process and introduce a novel metric, the ICR Score (Information Contribution to Residual Stream), which quantifies the contribution of modules to the hidden states' update. We empirically validate that the ICR Score is effective and reliable in distinguishing hallucinations. Building on these insights, we propose a hallucination detection method, the ICR Probe, which captures the cross-layer evolution of hidden states. Experimental results show that the ICR Probe achieves superior performance with significantly fewer parameters. Furthermore, ablation studies and case analyses offer deeper insights into the underlying mechanism of this method, improving its interpretability.