CausalGaze: Unveiling Hallucinations via Counterfactual Graph Intervention in Large Language Models

TL;DR

CausalGaze introduces a causal intervention framework using structural causal models to detect hallucinations in large language models, improving interpretability and accuracy.

Contribution

It pioneers active causal intervention with counterfactuals in LLM hallucination detection, surpassing static signal-based methods.

Findings

Achieves 3.3% AUROC improvement on TruthfulQA

Models LLM internal states as dynamic causal graphs

Effective across four datasets and three LLMs

Abstract

Despite the groundbreaking advancements made by large language models (LLMs), hallucination remains a critical bottleneck for their deployment in high-stakes domains. Existing classification-based methods mainly rely on static and passive signals from internal states, which often captures the noise and spurious correlations, while overlooking the underlying causal mechanisms. To address this limitation, we shift the paradigm from passive observation to active intervention by introducing CausalGaze, a novel hallucination detection framework based on structural causal models (SCMs). CausalGaze models LLMs' internal states as dynamic causal graphs and employs counterfactual interventions to disentangle causal reasoning paths from incidental noise, thereby enhancing model interpretability. Extensive experiments across four datasets and three widely used LLMs demonstrate the effectiveness of CausalGaze, especially achieving 3.3% improvement in AUROC on the TruthfulQA dataset compared to state-of-the-art baselines.