Cross-Layer Attention Probing for Fine-Grained Hallucination Detection

Malavika Suresh; Rahaf Aljundi; Ikechukwu Nkisi-Orji; Nirmalie Wiratunga

cs.CLSep 4, 2025

Cross-Layer Attention Probing for Fine-Grained Hallucination Detection

Authors: Malavika Suresh, Rahaf Aljundi, Ikechukwu Nkisi-Orji, Nirmalie Wiratunga

Organizations: Robert Gordon University, Aberdeen, United Kingdom · Toyota Motor Europe, Brussels, Belgium

Abstract

With the large-scale adoption of Large Language Models (LLMs) in various applications, there is a growing reliability concern due to their tendency to generate inaccurate text, i.e. hallucinations. In this work, we propose Cross-Layer Attention Probing (CLAP), a novel activation probing technique for hallucination detection, which processes the LLM activations across the entire residual stream as a joint sequence. Our empirical evaluations using five LLMs and three tasks show that CLAP improves hallucination detection compared to baselines on both greedy decoded responses as well as responses sampled at higher temperatures, thus enabling fine-grained detection, i.e. the ability to disambiguate hallucinations and non-hallucinations among different sampled responses to a given prompt. This allows us to propose a detect-then-mitigate strategy using CLAP to reduce hallucinations and improve LLM reliability compared to direct mitigation approaches. Finally, we show that CLAP maintains high reliability even when applied out-of-distribution. Code available at: https://github.com/itsmemala/CLAP

Figures & tables

Figure 1: Cross-Layer Attention Probing: For an LLM with

L

layers (transformer decoder blocks) and an input prompt

i

, activations

x_{i,l}

are collected at the output of each layer

l

when generating the last response token (EOS). The activations are first projected to a lower dimensional space through a learnable projection layer and the set of all projected activations forms a sequence of input tokens. A learnable CLS token is employed at the start of the sequence to extract features of hallucination. The sequence input is then fed to a transformer encoder block and the CLS embedding output is fed to a linear classifier layer for binary label prediction of hallucination/non hallucination decision.

LLM	PE	AH-g	LP-g	NLP-g	CLAP-g	AH-s	LP-s	NLP-s	CLAP-s
Greedy Test Responses
Llama 7B	69.0	64.0	77.7	77.2	78.1	69.2	75.3	75.6	77.8
Alpaca 7B	78.6	81.2	85.2	85.8	86.6	83.6	86.3	86.4	87.3
Vicuna 7B	80.4	78.0	88.4	88.2	88.5	85.7	88.8	88.8	88.7
Gemma 2B	62.8	73.5	70.2	71.2	72.7	74.1	70.1	71.6	72.7
Llama3.1-I 8B	69.7	66.6	66.1	66.5	68.1	67.6	67.8	68.3	69.0

Table 1: Hallucination detection performance of CLAP versus baselines, measured in AUC scores, best is indicated in bold, second best is underlined. ’-g’ and ’-s’ denote the use of greedy responses only and the use of both greedy and sampled responses for training, respectively. Results averaged across TQA, NQ and STR, across three random seeds.

	% Non-Hallucinations (NH) $\uparrow$					%Abs $\downarrow$		%Abs but NH $\downarrow$
Data	Def	Def+Abs *	Alt	+CLAP-I	+CLAP-II *	Def+Abs	+CLAP-II	Def+Abs	+CLAP-II
L-7B TQA	57.3	75.6	48.6	56.3	71.3	36.0	26.4	9.1	5.5
L-7B STR	60.0	69.9	64.0	65.9	67.6	45.0	7.2	21.4	3.9
A-7B TQA	34.2	68.3	33.8	41.9	68.1	62.7	44.2	8.7	6.3
A-7B STR	43.4	67.0	60.5	64.8	65.2	44.3	16.1	6.1	2.6
V-7B TQA	14.6	59.2	28.3	31.0	57.7	80.7	59.9	3.2	2.6

Table 2: Mitigating hallucinations with fine-grained detection using CLAP applied to DoLa. First block indicates the LLM and dataset. Second block shows the % of non-hallucinated responses for each of the five response generation strategies. Third and forth blocks show the % of responses abstained and the % of responses abstained but were non-hallucinated (NH), respectively. For results with +CLAP-I and +CLAP-II we report mean across three seeds. * denotes %NH among non-abstained responses.

Panel of Figure 2 — Figure 2: Transitions when replacing original greedy responses with responses using Alt and +CLAP-II strategies.

LLM	%gain-over-last	%gain-over-MA	%gain-over-MC	%gain-over-MV	%gain-over-SEP
Llama 7B	4.9	2.1	4.9	1.7	45.2
Alpaca 7B	0.9	1.4	1.1	1.3	18.0
Vicuna 7B	5.1	1.3	-1.8	5.2	35.6
Gemma 2B	6.3	3.6	1.4	3.8	25.1
Llama3.1-I 8B	2.4	-1.6	-2.2	5.9	15.1

Table 3: Out-of-distribution (OOD) hallucination detection performance of CLAP versus probe selection strategies and semantic entropy probing (SEP), measured in % gain on AUC. Results averaged across twenty OOD pairs derived from five datasets, across three random seeds, using greedy responses for train and test.

$n_{enc}$	$d_{model}$	TQA	NQ	STR	$n_{enc}$	$d_{model}$	TQA	NQ	STR
1	128	84.3 (0.2)	86.1 (0.8)	64.6 (0.7	2	128	84.3 (0.2)	86.0 (0.8)	64.5 (0.8)
1	256	84.3 (0.3)	86.1 (1.0)	64.2 (0.8)	2	256	84.3 (0.2)	86.2 (1.0)	64.7 (0.3)
1	512	84.2 (0.1)	86.0 (1.1)	64.8 (0.8)	2	512	84.3 (0.3)	86.1 (1.1)	65.1 (0.3)
1	1024	84.2 (0.4)	86.0 (0.7)	64.7 (0.9)	2	1024	84.3 (0.5)	86.3 (0.7)	65.2 (1.0)
1	2048	84.1 (0.4)	85.8 (1.0)	64.4 (0.5)	2	2048	84.4 (0.4)	86.0 (1.1)	65.7 (1.2)
1	4096*	83.8 (0.1)	86.0 (0.9)	63.6 (0.6)	2	4096*	83.9 (0.4)	86.0 (1.0)	66.7 (1.6)

Table 4: Analysis of hyper-parameter choices, using AUC on validation set, best is indicated in bold. * denotes no projection. Results on Llama 7B, averaged across three random seeds, using greedy responses for train and test.

Probe	$n_{enc}$	$d_{model}$	# Params	In Distribution		Out Of Distribution
				TQA	NQ	TQA-¿City	NQ-¿City
Maxpool	-	4096*	4K	77.4 (0.7)	82.9 (0.3)	50.9 (2.8)	51.8 (2.5)
Project + Concat	-	128	528K	82.7 (0.1)	87.0 (0.3)	56.7 (1.2)	54.1 (0.9)
Project + Concat	-	256	1M	82.6 (0.1)	86.9 (0.1)	56.8 (1.5)	54.4 (1.8)
Project + Concat	-	512	2.1M	82.8 (0.1)	86.9 (0.3)	57.0 (0.1)	53.2 (1.1)
Project + Concat	-	4096*	16.9M	83.1 (0.3)	87.1 (0.2)	57.5 (0.6)	53.9 (0.7)

Table 5: Ablating CLAP components and design, using AUC on test set, best is indicated in bold, second best is underlined. * denotes no projection, K denotes one thousand and M denotes one million. Results on Llama 7B, averaged across three random seeds, using greedy responses for train and test.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

		Greedy Responses				Sampled Responses
Model	Dataset	# Prompts		% Hallucinations		# Samples		% Hallucinations
		Train	Test	Train	Test	Train	Test	Train	Test
Llama-7B	TQA	5000	1800	45.9	42.7	10	10	79.1	76.9
	NQ	5000	1800	75.3	77.4	10	10	93.5	93.6
	STR	1832	458	39.8	40.0	8	8	68.3	67.2
Alpaca-7B	TQA	5000	1800	67.5	65.8	10	10	74.6	73.4

Appendix

Table 6: Statistics of the datasets used. Shows number of prompts, additional responses sampled per prompt and % of hallucinated responses in the train and test splits.

Model	Dataset	% Responses with a refusal phrase
		Train: Greedy + Sampled	Test: Greedy	Test: Sampled
Llama-7B	TQA	3.6	0.1	3.4
	NQ	4.5	0.0	5.2
	STR	4.4	5.0	3.9
Alpaca-7B	TQA	0.8	0.3	0.9
	NQ	1.2	0.9	1.2

Appendix

Table 7: Refusal rate of LLMs. Shows % of responses containing any one of a list of common refusal phrases.

Data	PE	AH-g	LP-g	NLP-g	CLAP-g	AH-s	LP-s	NLP-s	CLAP-s
Llama-7B
TQA	76.0	67.5 (0.1)	82.0 (0.1)	81.6 (0.4)	82.0 (0.0)	72.1 (0.0)	80.1 (0.2)	81.2 (0.4)	81.6 (0.2)
NQ	80.1	75.0 (0.1)	85.9 (0.4)	86.2 (0.2)	86.6 (0.3)	80.2 (0.7)	85.0 (0.2)	85.9 (0.1)	86.1 (0.4)
STR	50.9	49.5 (1.1)	65.0 (1.3)	63.9 (0.8)	65.7 (1.0)	55.4 (0.6)	61.0 (1.1)	59.9 (2.3)	65.6 (0.6)
Avg	69.0	64.0	77.7	77.2	78.1	69.2	75.3	75.6	77.8
Alpaca-7B

Appendix

Table 8: Hallucination detection performance of CLAP versus baselines on greedy responses to prompts, measured in AUC scores, best is indicated in bold, second best is underlined. ’-g’ and ’-s’ denote the use of greedy responses only and the use of both greedy and sampled responses for training, respectively.

Data	PE	AH-g	LP-g	NLP-g	CLAP-g	AH-s	LP-s	NLP-s	CLAP-s
Llama-7B
TQA	88.0	81.0 (0.5)	82.5 (0.5)	87.0 (0.7)	85.8 (1.8)	86.9 (0.1)	90.0 (0.7)	90.5 (0.0)	91.2 (0.1)
NQ	90.3	82.0 (0.3)	79.7 (2.9)	88.1 (0.3)	85.3 (0.6)	90.3 (0.2)	90.5 (0.8)	91.2 (0.7)	92.3 (0.2)
STR	74.6	71.5 (3.4)	65.2 (3.8)	57.4 (8.2)	51.6 (3.8)	82.1 (0.1)	85.8 (0.1)	85.3 (0.1)	86.4 (0.3)
Avg	84.3	78.2	75.8	77.5	74.3	86.5	88.8	89.0	89.9
Alpaca-7B

Appendix

Table 9: Hallucination detection performance of CLAP versus baselines on sampled responses to prompts, measured in AUC scores, best is indicated in bold, second best is underlined. ’-g’ and ’-s’ denote the use of greedy responses only and the use of both greedy and sampled responses for training, respectively.

Model	Data	LP-rs	NLP-rs	CLAP-rs	LP-pw	NLP-pw	CLAP-pw
Greedy Test Responses
L-7B	TQA	79.9 (1.0)	80.1 (0.2)	80.8 (0.9)	80.1 (0.2)	81.2 (0.4)	81.6 (0.2)
L-7B	NQ	84.3 (0.7)	85.0 (0.4)	84.7 (0.8)	85.0 (0.2)	85.9 (0.1)	86.1 (0.4)
L-7B	STR	60.6 (1.2)	59.9 (3.3)	66.5 (1.8)	61.0 (1.1)	59.9 (2.3)	65.6 (0.6)
	Avg	74.9	75.0	77.3	75.3	75.6	77.8
A-7B	TQA	89.4 (0.1)	88.9 (0.4)	89.1 (0.4)	89.3 (0.2)	89.3 (0.3)	89.3 (0.4)

Appendix

Table 10: Comparing effect of train-time batching strategy on hallucination detection performance, measured in AUC scores, best is indicated in bold, second best is underlined.

	% Non-Hallucinations $\uparrow$					%Abs $\downarrow$		%Abs but NH $\downarrow$
Data	Def	Def+Abs *	Alt	+CLAP-I	+CLAP-II *	Def+Abs	+CLAP-II	Def+Abs	+CLAP-II
L-7B TQA	57.3	75.6	22.8	51.8	71.8	36.0	29.7	9.1	6.9
L-7B STR	60.0	69.9	32.3	51.9	59.9	45.0	22.7	21.4	10.9
A-7B TQA	34.2	68.3	26.9	37.6	67.3	62.7	50.4	8.7	6.3
A-7B STR	43.4	67.0	39.7	48.2	61.4	44.3	31.6	6.1	3.0
V-7B TQA	14.6	59.2	6.6	15.6	53.7	80.7	72.6	3.2	2.8

Appendix

Table 11: Mitigating hallucinations with fine-grained detection using CLAP applied to Random Sampling. Second block shows the % of non-hallucinated responses for each of the five response generation strategies. Third and forth blocks show the % of responses abstained and the % of responses that were abstained but were non-hallucinated (NH), respectively. For results with +CLAP-I and +CLAP-II we report mean across three seeds. * denotes %non-hallucinations among non-abstained responses.

	% Non-Hallucinations $\uparrow$			H -¿ NH $\uparrow$			NH -¿ H $\downarrow$
LLM	+LP-I	+NLP-I	+CLAP-I	+LP-I	+NLP-I	+CLAP-I	+LP-I	+NLP-I	+CLAP-I
Llama 7B	60.2	60.0	61.1	55.8	45.8	57.6	34.4	26.8	32.8
Alpaca 7B	51.6	52.7	53.4	77.2	78.1	80.1	26.4	21.6	21.8

Appendix

Table 12: Hallucination Mitigation: CLAP-I versus Last Layer Probing when combined with DoLa, best is indicated in bold. Second block shows the % of non-hallucinated responses using each method. Third and fourth blocks show the ratio of % H-¿NH (or NH-¿H) replacements using LP-I/NLP-I/CLAP-I against % H-¿NH (or NH-¿H) replacements using Alt. Results averaged across TQA and STR, across three seeds.

	% Non-Hallucinations $\uparrow$			%Abs $\downarrow$			%Abs but NH $\downarrow$
LLM	+LP-II *	+NLP-II *	+CLAP-II *	+LP-II	+NLP-II	+CLAP-II	+LP-II	+NLP-II	+CLAP-II
Llama 7B	67.8	68.5	69.5	16.4	18.5	16.8	4.9	5.3	4.7
Alpaca 7B	66.6	64.9	66.7	33.1	29.2	30.2	5.6	4.2	4.5

Appendix

Table 13: Hallucination Mitigation: CLAP-II versus Last Layer Probing when combined with DoLa, best is indicated in bold. Second block shows the % of non-hallucinated responses using each method. Third and fourth blocks show the % of responses abstained and the % of responses that were abstained but were non-hallucinated (NH), respectively. * denotes %non-hallucinations among non-abstained responses. Results averaged across TQA and STR, across three seeds.

	H -¿ NH $\uparrow$			NH -¿ H $\downarrow$
LLM	+LP-II	+NLP-II	+CLAP-II	+LP-II	+NLP-II	+CLAP-II
Llama 7B	41.0	27.5	42.6	19.8	11.7	19.0
Alpaca 7B	49.2	57.0	60.4	3.8	4.2	4.2

Appendix

Table 14: Comparison of LP-II/NLP-II with DoLa versus CLAP-II with DoLa, best is indicated in bold. Shows the ratio of % H-¿NH (or NH-¿H) replacements using LP-II/NLP-II/CLAP-II against % H-¿NH (or NH-¿H) replacements using Alt. Results averageed across TQA and STR, across three seeds.

Probe	In Distribution	Out Of Distribution
	TQA-¿TQA	TQA-¿City
Att-Pool (MA)	77.7 (1.2)	72.3 (4.4)
Att-Pool-Ens	77.3(0.4)	66.4 (4.9)
CLAP	76.3 (0.3)	79.0 (3.9)

Appendix

Table 15: Comparing CLAP with token-wise attention-pooling, using AUC scores on test data, best is indicated in bold. Results on Gemma 2B, using greedy responses for train and test.

$n_{enc}$	$d_{model}$	TQA	NQ	STR	TQA	NQ	STR
		Alpaca 7B			Vicuna 7B
1	128	87.3 (0.4)	89.0 (1.2)	83.8 (0.2)	93.0 (0.9)	88.1 (0.1)	85.1 (1.4)
1	256	87.2 (0.5)	89.0 (1.0)	83.6 (0.2)	92.8 (0.8)	88.3 (0.3)	84.7 (1.2)
1	512	87.4 (0.5)	88.6 (1.2)	83.8 (0.6)	92.8 (0.9)	88.1 (0.3)	85.3 (1.3)
1	1024	87.3 (0.5)	88.5 (1.0)	83.7 (0.3)	93.0 (0.8)	88.1 (0.2)	85.1 (1.4)
1	2048	87.4 (0.5)	88.3 (1.0)	83.5 (0.3)	92.8 (0.9)	88.3 (0.2)	85.4 (1.4)

Appendix

Table 16: Analysis of hyper-parameter choices for CLAP, using AUC scores on validation data.

LLM Hallucination Detection

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

See all 208 papers

HALT: Hallucination Assessment via Log-probs as Time series
Feb 2, 2026Ahmad Shapiro, Karan Taneja, Ashok GoelLanguage Model Error Detection Hallucination Detection
Georgia Institute of Technology · Microsoft
Grad Detect: Gradient-Based Hallucination Detection in LLMs
Jun 23, 2026Anand Kamat, Daniel Blake, Brent M. WernessLLM Reliability LLM Hallucination Detection
Amazon
External Observers May See More Clearly: Cross-Model Span-Level Hallucination Detection in Large Language Models via Hidden State Probing
Oct 1, 2026Kingshuk Gupta, Davide BuscaldiHallucination Detection LLM Hallucination Detection
LIX, École Polytechnique · School of Electrical and Electronic Engineering, Nanyang Technological University · LIPN, Sorbonne Paris Nord

Feb 2, 2026J/K move · Enter open · S save

HALT: Hallucination Assessment via Log-probs as Time series

Ahmad Shapiro, Karan Taneja, Ashok Goel

Georgia Institute of Technology · Microsoft

Language Model Error Detection Hallucination Detection Hallucination in Language Models LLM Reliability LLM Hallucination Detection

Open paper AskPDF arXiv

Hallucinations remain a major obstacle for large language models (LLMs), especially in safety-critical domains. We present HALT (Hallucination Assessment via Log-probs as Time series), a lightweight hallucination detector that leverages only the top-20 token log-probabilities from LLM generations as a time series. HALT uses a gated recurrent unit model combined with entropy-based features to learn model calibration bias, providing an extremely efficient alternative to large encoders. Unlike white-box approaches, HALT does not require access to hidden states or attention maps, relying only on output log-probabilities. Unlike black-box approaches, it operates on log-probs rather than surface-form text, which enables stronger domain generalization and compatibility with proprietary LLMs without requiring access to internal weights. To benchmark performance, we introduce HUB (Hallucination detection Unified Benchmark), which consolidates prior datasets into ten capabilities covering both reasoning tasks (Algorithmic, Commonsense, Mathematical, Symbolic, Code Generation) and general purpose skills (Chat, Data-to-Text, Question Answering, Summarization, World Knowledge). While being 30x smaller, HALT outperforms Lettuce, a fine-tuned modernBERT-base encoder, achieving a 60x speedup gain on HUB. HALT and HUB together establish an effective framework for hallucination detection across diverse LLM capabilities.

Abstract

LLM Hallucination Detection

Explore similar work

HALT: Hallucination Assessment via Log-probs as Time series

Grad Detect: Gradient-Based Hallucination Detection in LLMs

External Observers May See More Clearly: Cross-Model Span-Level Hallucination Detection in Large Language Models via Hidden State Probing