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

Linggang Kong; Lei Wu; Yunlong Zhang; Xiaofeng Zhong; Zhen Wang; Yongjie Wang; Yao Pan

cs.LGApr 13, 2026

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

Authors: Linggang Kong, Lei Wu, Yunlong Zhang, Xiaofeng Zhong, Zhen Wang, Yongjie Wang, Yao Pan

Organizations: College of Electronic Engineering, National University of Defense Technology · Anhui Province Key Laboratory of Cyberspace Security Situation Awareness and Evaluation · Institute of Computer Application, China Academy of Engineering Physics

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.

Figures & tables

Figure 1: An question-answering example with noise and spurious correlations in graph structure. The nodes and edges are from the raw hidden states and attention maps, respectively.

Figure 2: Overall framework of the proposed CausalGaze. We first employ the gradient-guided counterfactual intervention to obtain the refined edges, then the hidden states and the refined edges are together passed to the GNN-based detection module for the final hallucination detection result.

Figure 3: Visualization of the gradients for the example in Figure 1 .The noise connection from token ‘ # ’ to token ‘ ## ’ and spurious connection from token ‘ We ’ to token ‘California ’ have a near-zero value.

Panel of Figure 4 — Figure 4: The comparison of graph structures before and after the intervention for the example in Figure 1 . (a) The edges in raw graph are from the observation attention maps. (b) The edges in refined graph are from the actively intervention attention maps.

LLMs	Methods	TruthfulQA		TriviaQA		SciQ		HaluEval
LLMs	Methods	AUROC	F1-Score	AUROC	F1-Score	AUROC	F1-Score	AUROC	F1-Score
Llama2-7B	SelfCheckGPT	0.5295	0.5345	0.7322	0.7296	0.6790	0.6830	0.6670	0.6591
	EigenScore	0.5193	0.5286	0.7398	0.7443	0.5960	0.6008	0.5960	0.5880
	Perplexity	0.5677	0.6034	0.7213	0.6995	0.5260	0.5256	0.5260	0.5020
	LN-Entropy	0.6151	0.6187	0.7091	0.7332	0.5760	0.5826	0.6678	0.6544
	Semantic Entropy	0.6217	0.6145	0.7321	0.7209	0.6820	0.6797	0.6820	0.6865

Table 1: Main results of AUROC and F1-Score compared with different competitive baselines over diverse LLMs on TruthfulQA, TriviaQA, SciQ and HaluEval datasets. The best results are highlighted in bold.

Figure 5: Cross-dataset generalization analysis for CausalGaze and HaluGNN. Each subplot displays the F1-Score when the model is trained on the row dataset and tested on the column dataset, with values annotated in each cell.

Projection Layer	Dimension	TruthfulQA		TriviaQA		SciQ		Halueval
Projection Layer	Dimension	AUROC	F1-Score	AUROC	F1-Score	AUROC	F1-Score	AUROC	F1-Score
w/o	4096	0.6069	0.6469	0.5898	0.5318	0.5889	0.6559	0.5530	0.6676
with	64	0.8712	0.7818	0.8719	0.7809	0.8301	0.7773	0.8989	0.6322
	128	0.8805	0.8189	0.8614	0.7786	0.8390	0.7701	0.9280	0.8001
	256	0.8413	0.7822	0.8840	0.8011	0.8324	0.7462	0.8837	0.7093
	512	0.8807	0.8018	0.8849	0.7956	0.8267	0.7333	0.8168	0.7698

Table 2: The contribution of the Projection Layer and its dimentions on Llama2-7B across different datasets. The best results are highlighted in bold.

Figure 6: The impact of causal graphs from different layers of Llama2-7B on the detection performance.

Model	Variant	TruthfulQA	TriviaQA	SciQ	HaluEval
Llama2-7B	CausalGaze	0.8189	0.7786	0.7701	0.8001
	w/o Gradient	0.7720	0.7763	0.7549	0.6316
	Random Gradient	0.7811	0.7656	0.7319	0.6630
	MLP( $A$ )	0.7793	0.7674	0.7540	0.7241
Qwen2-7B	CausalGaze	0.8071	0.8280	0.8283	0.8360
	w/o Gradient	0.7660	0.8240	0.8400	0.8033

Table 3: The performance comparison of CausalGaze with w/o Gradient, Random Gradient and A learnable gating mechanism based solely on the attention maps (

A

) (i.e., MLP(

A

)).

Panel of Figure 7 — Figure 7: The token-level and fine-grained interpretability analysis of the detection result for the hallucinated and factual responses of the same question in Figure 1 .

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Methods	TruthfulQA	TriviaQA	SciQ	HaluEval
Lookback Lens	0.603 / 0.633	0.771 / 0.781	0.726 / 0.731	0.752 /0 .795
Factoscope	0.552 / 0.584	0.714 / 0.760	0.682 / 0.702	0.742 / 0.788
TAD	0.804 / 0.832	0.844 / 0.849	0.815 / 0.833	0.874 / 0.856
CausalGaze	0.881 / 0.852	0.861 / 0.895	0.839 / 0.855	0.928 / 0.938

Appendix

Table 4: AUROC / PR-AUC Performance of different methods on Llama2-7B.

LLMs	Projection Layer	Dimension	TruthfulQA		TriviaQA		SciQ		Halueval
LLMs	Projection Layer	Dimension	AUROC	F1-Score	AUROC	F1-Score	AUROC	F1-Score	AUROC	F1-Score
Qwen2-7B	w/o	3584	0.5925	0.5663	0.6043	0.5728	0.5964	0.6459	0.5620	0.6588
	with	64	0.8330	0.7845	0.8764	0.7890	0.8976	0.8423	0.8946	0.7892
		128	0.8680	0.8071	0.9106	0.8280	0.9328	0.8283	0.9220	0.8360
		256	0.8357	0.7786	0.8873	0.8130	0.8741	0.7740	0.8863	0.7538
		512	0.8588	0.7943	0.8743	0.7924	0.8661	0.7548	0.8750	0.7630

Appendix

Table 5: The contribution of the Projection Layer and its dimentions on Qwen2-7B and Mistral-7B across different datasets. The best results are highlighted in bold.

LLMs	Layer Number	TruthfulQA		TriviaQA		SciQ		Halueval
LLMs	Layer Number	AUROC	F1-Score	AUROC	F1-Score	AUROC	F1-Score	AUROC	F1-Score
Qwen2-7B	16	0.8718	0.7836	0.9005	0.8194	0.9114	0.8445	0.9260	0.8004
	20	0.8776	0.7893	0.9125	0.8304	0.9268	0.8675	0.9335	0.8421
	24	0.8892	0.7747	0.8994	0.7961	0.8993	0.8593	0.9154	0.8266
	28	0.8757	0.7694	0.8873	0.7868	0.8858	0.8495	0.9255	0.8273
	32	0.8680	0.8071	0.9106	0.8280	0.9328	0.8283	0.9220	0.8360

Appendix

Table 6: The impact of causal graphs from different layers of Qwen2-7B and Mistral-7B on the detection performance.

Figure 8: Cross-dataset generalization analysis for CausalGaze on Qwen2-7B and Mistral-7B. Each subplot displays the F1-Score when the model is trained on the row dataset and tested on the column dataset, with values annotated in each cell.

Method	Input memory	Model parameters	Inference latency
SAPLMA	608KB	109K	$\sim$ 0.307ms
HaluGNN	614KB	213K	$\sim$ 0.903ms
SelfCheckGPT	608KB	-	$\sim$ 10.32s
CausalGaze	614KB	575K	$\sim$ 1.731ms

Appendix

Table 7: Computation overhead comparison across different baselines.

LLM Hallucination Detection

Large language model hallucination detection identifies unsupported, incorrect, or fabricated claims in model outputs and traces.

See all 208 papers

HalluSAE: Detecting Hallucinations in Large Language Models via Sparse Auto-Encoders
Apr 6, 2026Boshui Chen, Zhaoxin Fan, Ke Wang +5LLM Hallucination Detection
Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, School of Artificial Intelligence, Beihang University, Beijing, China. · State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China. · Beijing Academy of Blockchain and Edge Computing, Beijing, China. +1
Unmasking Hallucinations: A Causal Graph-Attention Perspective on Factual Reliability in Large Language Models
Apr 5, 2026Sailesh kiran kurra, Shiek Ruksana, Vishal BorusuLLM Hallucination Mitigation Factual Knowledge in Language Models
Amazon,Texas, United States of America · Vasavi College of Engineering, Telangana,India
HART: Data-Driven Hallucination Attribution and Evidence-Based Tracing for Large Language Models
Mar 6, 2026Shize Liang, Hongzhi WangHallucination in Language Models LLM Interpretability
Faculty of Computing, Harbin Institute of Technology, Harbin, China

Apr 6, 2026J/K move · Enter open · S save

HalluSAE: Detecting Hallucinations in Large Language Models via Sparse Auto-Encoders

Boshui Chen, Zhaoxin Fan, Ke Wang, Zhiying Leng+4

Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, School of Artificial Intelligence, Beihang University, Beijing, China. · State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China. · Beijing Academy of Blockchain and Edge Computing, Beijing, China.+1

LLM Hallucination Detection

Open paper AskPDF arXiv

Large Language Models (LLMs) are powerful and widely adopted, but their practical impact is limited by the well-known hallucination phenomenon. While recent hallucination detection methods have made notable progress, we find most of them overlook the dynamic nature and underlying mechanisms of it. To address this gap, we propose HalluSAE, a phase transition-inspired framework that models hallucination as a critical shift in the model's latent dynamics. By modeling the generation process as a trajectory through a potential energy landscape, HalluSAE identifies critical transition zones and attributes factual errors to specific high-energy sparse features. Our approach consists of three stages: (1) Potential Energy Empowered Phase Zone Localization via sparse autoencoders and a geometric potential energy metric; (2) Hallucination-related Sparse Feature Attribution using contrastive logit attribution; and (3) Probing-based Causal Hallucination Detection through linear probes on disentangled features. Extensive experiments on Gemma-2-9B demonstrate that HalluSAE achieves state-of-the-art hallucination detection performance.

Abstract

LLM Hallucination Detection

Explore similar work

HalluSAE: Detecting Hallucinations in Large Language Models via Sparse Auto-Encoders

Unmasking Hallucinations: A Causal Graph-Attention Perspective on Factual Reliability in Large Language Models

HART: Data-Driven Hallucination Attribution and Evidence-Based Tracing for Large Language Models