Detecting Contextual Hallucinations in LLMs with Frequency-Aware Attention

Siya Qi; Yudong Chen; Runcong Zhao; Qinglin Zhu; Zhanghao Hu; Wei Liu; Yulan He; Zheng Yuan

cs.CLFeb 20, 2026

Detecting Contextual Hallucinations in LLMs with Frequency-Aware Attention

Authors: Siya Qi, Yudong Chen, Runcong Zhao, Qinglin Zhu, Zhanghao Hu, Wei Liu, Yulan He, Zheng Yuan,

Organizations: Department of Informatics, King’s College London, UK · Department of Statistics, University of Warwick, UK · The Alan Turing Institute, UK · School of Computer Science, The University of Sheffield, UK

Abstract

Hallucination detection is critical for ensuring the reliability of large language models (LLMs) in context-based generation. Prior work has explored intrinsic signals available during generation, among which attention offers a direct view of grounding behavior. However, existing approaches typically rely on coarse summaries that fail to capture fine-grained instabilities in attention. Inspired by signal processing, we introduce a frequency-aware perspective on attention by analyzing its variation during generation. We model attention distributions as discrete signals and extract high-frequency components that reflect rapid local changes in attention. Our analysis reveals that hallucinated tokens are associated with high-frequency attention energy, reflecting fragmented and unstable grounding behavior. Based on this insight, we develop a lightweight hallucination detector using high-frequency attention features. Experiments on the RAGTruth and HalluRAG benchmarks show that our approach achieves performance gains over verification-based, internal-representation-based, and attention-based methods across models and tasks.

Figures & tables

Figure 1: Attention weights over context and previously generated tokens for a grounded token (blue, “rainy”) and a hallucinated token (red, “December”) in a context-based QA example.

Figure 2: Three hypotheses for identifying hallucination tokens from incoming attention patterns. We illustrate three representative assumptions for distinguishing well-grounded tokens (✓) from potential hallucinations (✗) based on the incoming attention to the next generated token.

Figure 3: Overview of frequency-aware attention modeling for hallucination detection. Attention weights are extracted from each layer and head ( $L$ layers and $H$ heads in total), treated as token-level signals, and decomposed using high-pass filtering to isolate high-frequency variations ( $\mathcal{F}_{\text{high}}$ ), whose energy is aggregated for hallucination detection.

Model	Method	RT-QA		RT-D2T		RT-Summ		HalluRAG		Overall Avg.
Model	Method	F	AUROC	F	AUROC	F	AUROC	F	AUROC	Avg-F	Avg-A
LLaMA-7B	SelfcheckGPT	0.6289	0.6942	0.6552	0.8026	0.6349	0.6674	0.5388	0.5963	0.6145	0.6901
	RefChecker	0.5914	0.5865	0.5713	0.6349	0.6059	0.6080	0.5819	0.5722	0.5876	0.6004
	EigenScore	0.4979	0.5253	0.5256	0.5297	0.5065	0.4989	0.5364	0.6127	0.5166	0.5416
	Redeep	0.5972	0.6364	0.4791	0.3960	0.5897	0.5760	0.5912	0.6490	0.5643	0.5643
	Lookback-lens	0.6930	0.8482	0.6175	0.8442	0.5328	0.7156	0.6266	0.7405	0.6175	0.7871

Table 1: Comparison of different methods on RAGTruth and HalluRAG. Wavelet-high and Fourier-high denote leveraging the high-frequency components obtained via DWT and DFT, respectively. Bold and underlined values indicate the best and second-best performance within each model group.

Model	Method	RagTruth-Avg		HalluRAG
Model	Method	F	AUROC	F	AUROC
LLaMA-7B	Lookback-lens	0.6773	0.7884	0.6623	0.7856
	Laplacian	0.6747	0.7877	0.6775	0.7754
	Wavelet-High	0.6911	0.8117	0.6812	0.7928
	Fourier-high	0.7003	0.8412	0.6866	0.8100
LLaMA-13B	Lookback-lens	0.6781	0.8183	0.7096	0.8505

Table 2: Performance comparison on RagTruth-Avg and HalluRAG for sliding-window=8. Within each model, the best result is highlighted in bold , and the second-best is underlined .

Figure 4: Comparing full-, low-, and high-pass Fourier attention features. Average AUROC across models under token- and span-level evaluation settings.

Figure 5: Layer-wise importance of high-frequency Fourier-high attention features for LLaMA-7B.

RagTruth-Avg
Method	Original	Top- $k$ heads
Method	Original	k = 100	50	10
Laplacian	0.8003	0.7824	0.7666	0.7131
Wavelet-high	0.8087	0.7900	0.7708	0.7193
Fourier-high	0.8205	0.7915	0.7684	0.6995
HalluRAG

Table 3: AUROC Results for LLaMA-7B: Original vs. Top-

k

head only performance.

Figure 6: Average absolute classifier importance assigned to context-attention and generated-attention features over all examined models and datasets.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Task	Train	Val	Test	Prompt Len.	Response Len.	Halluc. Ratio
RagTruth	QA	839	–	150	439.5	208.1	0.1045
	D2T	883	–	150	892.8	225.4	0.0768
	Summ	793	–	150	840.1	153.4	0.0527
HalluRAG	QA	756	162	162	649.3	44.4	0.1530

Appendix

Table A1 : Statistics of datasets used in our experiments. Length and ratio are calculated on the token level.

Method	Target	Token-Level (Source)			Span-Level (Source)
		QA	D2T	Summ	QA	D2T	Summ
Lookback-lens	QA	0.8482	0.7839	0.7741	0.8467	0.8140	0.7985
	D2T	0.6340	0.8442	0.7211	0.6334	0.8551	0.7275
	Summ	0.6618	0.6400	0.7156	0.6563	0.6496	0.6635
Laplacian	QA	0.8449	0.7928	0.7881	0.8365	0.7898	0.7857
	D2T	0.6438	0.8519	0.6809	0.6147	0.8646	0.6264

Appendix

Table A2: Cross-domain evaluation. Models are trained on the column domain (Source) and evaluated on the row domain (Target). Bold values indicate in-domain performance (diagonal, Target=Source). Darker shading corresponds to higher cross-domain performance.

Model / Method	RT-QA		RT-D2T		RT-Summ		HalluRAG		Overall Avg.
Model / Method	F	AUROC	F	AUROC	F	AUROC	F	AUROC	Avg-F	Avg-A
LLaMA-7B
Lookback-lens	0.7218	0.8467	0.7249	0.8551	0.5852	0.6635	0.6623	0.7856	0.6736	0.7877
Attn-variance	0.5077	0.6979	0.4714	0.5886	0.4840	0.6348	0.4446	0.5209	0.4769	0.6106
Attn-entropy	0.7292	0.8502	0.7060	0.8457	0.5489	0.6637	0.6282	0.6909	0.6531	0.7626
Laplacian	0.7074	0.8365	0.7334	0.8646	0.5833	0.6619	0.6775	0.7754	0.6754	0.7846

Appendix

Table A3: Full performance comparison on RagTruth and HalluRAG with chunk size set to 8. For each model, the best result is highlighted in bold , and the second-best result is underlined .

Panel of Figure 13 — Figure A1 : Ablation study on different frequency cutoffs at token and span levels on hallucination detection performance. The top row (a-d) shows results at the token level, while the bottom row (e-h) shows results at the span level.

Model/Level	Padding	RT-QA		RT-D2T		RT-Summ		HalluRAG		Overall Avg.
		F	AUROC	F	AUROC	F	AUROC	F	AUROC	Avg-F	Avg-A
LLaMA-7B
level1	zero	0.7194	0.8526	0.6898	0.8569	0.5929	0.7165	0.6384	0.7550	0.6601	0.7953
	period	0.7146	0.8514	0.6887	0.8517	0.5799	0.7091	0.6360	0.7449	0.6548	0.7893
	symm	0.7215	0.8557	0.6875	0.8545	0.5880	0.7182	0.6353	0.7521	0.6581	0.7951
level2	zero	0.7157	0.8522	0.6893	0.8531	0.5822	0.7088	0.6359	0.7464	0.6558	0.7901

Appendix

Table A4: Performance comparison on token level using Wavelet-high across levels and padding methods. Best results within each model are in bold ; second-best are underlined .

Model/Level	Padding	RT-QA		RT-D2T		RT-Summ		HalluRAG		Overall Avg.
		F	AUROC	F	AUROC	F	AUROC	F	AUROC	Avg-F	Avg-A
LLaMA-7B
level1	zero	0.7313	0.8611	0.7479	0.8785	0.6020	0.7169	0.6846	0.8047	0.6915	0.8153
	period	0.7271	0.8572	0.7426	0.8748	0.5973	0.7008	0.6786	0.7963	0.6864	0.8073
	symm	0.7331	0.8680	0.7478	0.8821	0.6045	0.7199	0.6812	0.7928	0.6917	0.8157
level2	zero	0.7277	0.8594	0.7461	0.8760	0.5996	0.6997	0.6800	0.7953	0.6884	0.8076

Appendix

Table A5: Performance comparison when chunk size is 8 using Wavelet-high across levels and padding methods. Best results within each model are in bold ; second-best are underlined .

Figure A2 : Full results of comparing Full-, Low-, and High-Pass Fourier attention features. Average AUROC across models under token- and span-level evaluation settings.

Panel of Figure 17 — Figure A3 : Full results for layer-wise importance. Solid and dashed lines correspond to token-level and span-level detection, respectively. The shaded region indicates the standard deviation of head-level importance within each layer.

Table A6: Original vs. Top-

k

head only Performance.

Model	Setting	Operator	RagTruth-Avg		HalluRAG
			F	AUROC	F	AUROC
LLaMA-7B	Original	Laplacian	0.6588	0.8003	0.6370	0.7429
		Wavelet-high	0.6673	0.8087	0.6384	0.7550
		Fourier-high	0.6685	0.8205	0.6478	0.7629
	Context-only	Laplacian	0.6504	0.8014	0.6340	0.7396
		Wavelet-high	0.6541	0.8074	0.6402	0.7479

Appendix

Table A7 : Ablation study comparing context-only and generated-only attention features. Results are reported across spectral operators and models.

Panel of Figure 20 — Figure A4 : Raw attention signal visualizations. Each subfigure compares raw attention distributions for a hallucinated token (left) and a non-hallucinated token (right). Red and yellow curves denote attention over context and generated tokens, respectively, when generating a hallucinated token, while blue and green curves denote attention over context and generated tokens for a non-hallucinated token.

Hallucination Detection

Hallucination detection identifies unsupported, fabricated, or evidence-inconsistent content in language and multimodal model outputs.

See all 74 papers

RAGognizer: Hallucination-Aware Fine-Tuning via Detection Head Integration
Apr 17, 2026Fabian Ridder, Laurin Lessel, Malte SchillingHallucination in Language Models RAG Robustness
Computer Science Department, University of Münster, Münster, Germany
Detecting Hallucination in LLMs: Tracing the Topological Signatures of Impaired Context Sharing
Sep 17, 2026Amir Jalilifard, Anderson Rocha, Eric Wong +1LLM Hallucination Detection Hallucination in Language Models
Institute of Computing, Universidade Estadual de Campinas (UNICAMP), Campinas, SP, Brazil · Department of Computer and Information Science, University of Pennsylvania, Philadelphia, PA, USA
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

Apr 17, 2026J/K move · Enter open · S save

RAGognizer: Hallucination-Aware Fine-Tuning via Detection Head Integration

Fabian Ridder, Laurin Lessel, Malte Schilling

Computer Science Department, University of Münster, Münster, Germany

Hallucination in Language Models RAG Robustness LLM Hallucination Mitigation Retrieval-Augmented Generation Language Modeling LLM Hallucination Detection+1

Open paper AskPDF arXiv

Retrieval-Augmented Generation (RAG) is widely used to augment the input to Large Language Models (LLMs) with external information, such as recent or domain-specific knowledge. Nonetheless, current models still produce closed-domain hallucinations and generate content that is unsupported by the retrieved context. Current detection approaches typically treat hallucination as a post-hoc problem, relying on black-box consistency checks or probes over frozen internal representations. In this work, we demonstrate that hallucination detection based on internal state representation can also serve as a direct training signal. We introduce RAGognize, a dataset of naturally occurring closed-domain hallucinations with token-level annotations, and RAGognizer, a hallucination-aware fine-tuning approach that integrates a lightweight detection head into an LLM, allowing for the joint optimization of language modeling and hallucination detection. This joint objective forces the model to improve the separability of its internal states regarding hallucinations while simultaneously learning to generate well-formed and meaningful responses. Across multiple benchmarks, RAGognizer achieves state-of-the-art token-level hallucination detection while substantially reducing hallucination rates during generation, without degrading language quality or relevance.

Abstract

Hallucination Detection

Explore similar work

RAGognizer: Hallucination-Aware Fine-Tuning via Detection Head Integration

Detecting Hallucination in LLMs: Tracing the Topological Signatures of Impaired Context Sharing

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