CausalGaze: Unveiling Hallucinations via Counterfactual Graph Intervention in Large Language Models
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866914550807265280 |
|---|---|
| author | Kong, Linggang Wu, Lei Zhang, Yunlong Zhong, Xiaofeng Wang, Zhen Wang, Yongjie Pan, Yao |
| author_facet | Kong, Linggang Wu, Lei Zhang, Yunlong Zhong, Xiaofeng Wang, Zhen Wang, Yongjie Pan, Yao |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_11087 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | CausalGaze: Unveiling Hallucinations via Counterfactual Graph Intervention in Large Language Models Kong, Linggang Wu, Lei Zhang, Yunlong Zhong, Xiaofeng Wang, Zhen Wang, Yongjie Pan, Yao Machine Learning 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. |
| title | CausalGaze: Unveiling Hallucinations via Counterfactual Graph Intervention in Large Language Models |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2604.11087 |