DEX-AR: A Dynamic Explainability Method for Autoregressive Vision-Language Models

Walid Bousselham; Angie Boggust; Hendrik Strobelt; Hilde Kuehne

cs.CVMar 6, 2026

DEX-AR: A Dynamic Explainability Method for Autoregressive Vision-Language Models

Authors: Walid Bousselham, Angie Boggust, Hendrik Strobelt, Hilde Kuehne

Organizations: Tuebingen AI Center · University of Tuebingen · MIT CSAIL · MIT-IBM Watson AI Lab ·

Abstract

As Vision-Language Models (VLMs) become increasingly sophisticated and widely used, it becomes more and more crucial to understand their decision-making process. Traditional explainability methods, designed for classification tasks, struggle with modern autoregressive VLMs due to their complex token-by-token generation process and intricate interactions between visual and textual modalities. We present DEX-AR (Dynamic Explainability for AutoRegressive models), a novel explainability method designed to address these challenges by generating both per-token and sequence-level 2D heatmaps highlighting image regions crucial for the model's textual responses. The proposed method offers to interpret autoregressive VLMs-including varying importance of layers and generated tokens-by computing layer-wise gradients with respect to attention maps during the token-by-token generation process. DEX-AR introduces two key innovations: a dynamic head filtering mechanism that identifies attention heads focused on visual information, and a sequence-level filtering approach that aggregates per-token explanations while distinguishing between visually-grounded and purely linguistic tokens. Our evaluation on ImageNet, VQAv2, and PascalVOC, shows a consistent improvement in both perturbation-based metrics, using a novel normalized perplexity measure, as well as segmentation-based metrics.

Figures & tables

Figure 1: Example of token-level and sentence-level attribution maps for Vision-Language Models (VLMs). Given an input image and prompt, DEX-AR produces per-token heatmaps highlighting relevant image regions for each generated word. These are then aggregated into a final sentence-level heatmap using token-specific weighting scores that reflect visual relevance.

Figure 2: Architecture overview of DEX-AR. At each layer

l

, head

i

and generation step

t

, gradients of attention maps are computed and weighted based on their relative focus on visual versus textual tokens to produce attribution maps.

Model	Method	ImageNet		VQAv2		( $\downarrow)$ sec.
Model	Method	Pos( $\uparrow$ )	Neg.( $\downarrow$ )	Pos( $\uparrow$ )	Neg.( $\downarrow$ )	/img.
LlaVA-1.5	Attention	2.17	1.01	0.88	0.89	0.45
	GradCAM	1.43	1.1	0.91	0.77	0.66
	CheferCAM	2.06	1.03	0.93	0.78	6.20
	Attn $\times$ Grad	1.94	1.00	0.90	0.80	0.65
	Int.Grad	1.63	1.52	0.91	0.80	8.20

Table 1: Perturbation-based evaluation across different VLM architectures (Decoder-only, Encoder-Decoder, Prefix-Decoder) on ImageNet and VQAv2. Values represent Area Under the perplexity Curve (AUC) for positive (

\uparrow

) and negative (

\downarrow

) perturbations.

Model	Method	soft-IoU( $\uparrow$ )	IoU( $\uparrow$ )	EPG( $\uparrow$ )
LlaVA-1.5	Attention	2.00	19.10	16.00
	Rollout	2.54	19.59	11.41
	GradCAM	10.20	28.90	19.30
	CheferCAM	1.60	21.01	17.10
	Attn $\times$ Grad	5.10	24.20	26.60
	DEX-AR	17.70	36.34	27.75

Table 2: Segmentation Performance on PascalVOC. Comparison of different methods on four VLM architectures using soft-IoU and IoU, and EPG (Energy Pointing Game).

	LlaVA		BakLlaVA
$k_{\%}$	SNR $(\uparrow)$	MSE $(\downarrow)$	SNR $(\uparrow)$	MSE $(\downarrow)$
-	1.64	0.33	5.27	0.15
max	3.64	0.22	6.02	0.14
0.05	3.35	0.24	4.13	0.17
0.1	3.09	0.25	4.13	0.19
0.2	2.45	0.28	3.18	0.22

Table 3: Head Filtering Ablation: Analysis of different filtering thresholds (

k_{\%}

) measuring Signal-to-Noise Ratio (SNR) and Mean Squared Error (MSE) on PascalVOC.

Filtering		SNR $(\uparrow)$	MSE $(\downarrow)$	EPG $(\uparrow)$
Head	Filler	SNR $(\uparrow)$	MSE $(\downarrow)$	EPG $(\uparrow)$
✗	✗	9.16	0.13	44.96
✓	✗	16.84	0.10	63.38
✗	✓	89.29	0.11	92.50
✓	✓	96.12	0.12	95.04

Table 4: Filler words Filtering Ablation: Analysis of the impact of the different filtering strategies on the filtering of the filler words produced by the VLM on PascalVOC using LlaVA-1.5-7B.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

ReLU	Interm. feat.	LLaVA	BakLLaVA	PaliGemma	Florence-2
✗	✓	31.56	35.09	22.90	30.44
✓	✗	33.05	32.87	22.85	26.91
✗	✗	24.54	27.80	22.83	27.01
✓	✓	36.34	35.85	23.55	32.48

Appendix

Table 5: Ablation study results showing IoU scores on PascalVOC across different model architectures. ”ReLU” indicates the use of ReLU activation on attention gradients, and ”Interm. feat.” indicates the use of intermediate layer features to compute the gradient.

Method	POS $\uparrow$	NEG $\downarrow$	IC $\uparrow$	Insert. $\downarrow$	Del. $\uparrow$	sec./img. $\downarrow$
Rollout	6.13	3.12	0.69	0.14	0.21	0.51
GradCAM	7.49	4.39	0.64	0.20	0.28	0.66
CheferCAM	11.15	4.07	0.64	0.19	0.36	6.20
Attn x CAM	12.60	3.74	0.65	0.18	0.42	0.65
Integrated Grad.	13.50	12.77	0.44	0.32	0.33	3.25
RISE	6.39	3.87	0.65	0.21	0.28	11.80

Appendix

Table 6: Comprehensive evaluation of attribution methods on ImageNet using BakLlaVA-7B. POS and NEG refer to positive and negative perturbation metrics from the main paper, IC represents the PIC metric, Insert. and Del. are the insertion and deletion metrics, and sec./img. shows computational efficiency.

Figure 4: PascalVOC-QA: example of the dataset used to evaluate the quality of the filtering.

Figure 5: Qualitative examples of the heatmap generated by DEX-AR for different VLMs with and without filtering.

Figure 6: Qualitative examples of failures cases of DEX-AR for different VLMs.

Figure 7: Qualitative evaluation on a complex, cluttered scene. DEX-AR accurately localizes distinct objects (e.g., ”hat”, ”clock”) and abstract concepts (”vintage”) despite the dense environment. Notably, the method distinguishes between visually grounded tokens (e.g., ”suit”) and purely linguistic completions (e.g., ”case”), effectively suppressing the latter.

Figure 8: Qualitative comparison between TAM and DEX-AR. Given the prompt “A gray cat is sitting next to an orange cat [EOS]”, we visualize the attribution maps for the first occurrence of “cat” (referring to the gray cat) and the second (referring to the orange cat). DEX-AR produces sharp, instance-specific localizations that accurately reflect the distinct visual grounding of each token. In contrast, TAM produces more diffuse maps, relying on static feature projections and heuristic filtering.

Model	Method	Positive (AUC) $\uparrow$	Negative (AUC) $\downarrow$
LLaVA-1.5	TAM	1.82	1.03
	DEX-AR (Ours)	2.31	0.96

Appendix

Table 7: Perturbation-based evaluation on ImageNet comparing DEX-AR and TAM.

Figure 9: Visual examples of the ImageNet-C corruptions used for robustness evaluation at severity level 5.

Filter Setting	Pixelate	Fog	Brightness	Zoom Blur	Gauss. Noise	Motion Blur
max (Ours)	125.90	153.53	132.02	108.97	62.72	97.72
avg	7.89	9.43	9.44	5.43	1.48	6.17
top- $k_{5\%}$	118.83	140.65	135.01	103.62	47.79	97.54
top- $k_{10\%}$	103.09	128.20	122.95	90.05	30.46	76.43
top- $k_{30\%}$	96.10	149.53	137.96	62.25	-15.02	48.10
top- $k_{50\%}$	-1.50	7.47	10.07	-10.22	-16.82	-9.03

Appendix

Table 8: Robustness evaluation on ImageNet-C (Severity 5). We report the Signal-to-Noise Ratio (SNR) for different filtering strategies across six corruption types. Higher values indicate better alignment with the target object despite the corruption. Best results are marked in bold .

Figure 10: Robustness to Registers. Side-by-side comparison of raw attention maps (middle) and DEX-AR heatmaps (right). Raw attention maps frequently suffer from ”register” artifacts which correspond to high-norm tokens in uninformative background regions (highlighted by pink circles). DEX-AR effectively suppresses these artifacts, demonstrating that the gradient-based importance score correctly identifies them as irrelevant to the specific token prediction, focusing instead on the semantic subjects (e.g., the cat, the horse rider).

VLM Interpretability

Vision-language model interpretability explains how image and language representations shape predictions, evidence use, and multimodal reasoning.

See all 268 papers

Angular Supervision for Visual Object Grounding in Autoregressive Vision-Language Models
Oct 23, 2026Parsa Esmaeilkhani, Longin Jan LateckiVision-Language Grounding VLM Interpretability
Temple University, Philadelphia, USA
Language Models Can Explain Visual Features via Steering
Mar 23, 2026Javier Ferrando, Enrique Lopez-Cuena, Pablo Agustin Martin-Torres +3VLM Interpretability
Barcelona Supercomputing Center
Attention-Guided Saliency Maps for Interpreting Visualization Literacy in VLMs
Jul 17, 2026Maeve Hutchinson, Abderrahmane Wassim Mehdaoui, Pranava MadhyasthaData Visualization Visual Attention
City St George’s · City St George’s, University of London · The Alan Turing Institute

Oct 23, 2026J/K move · Enter open · S save

Angular Supervision for Visual Object Grounding in Autoregressive Vision-Language Models

Parsa Esmaeilkhani, Longin Jan Latecki

Temple University, Philadelphia, USA

Vision-Language Grounding VLM Interpretability Representation Geometry in VLMs Referring Expression Segmentation

Open paper AskPDF arXiv

Autoregressive vision-language models (VLMs) struggle to ground generated text in the correct image regions, and existing explainability methods, such as attention maps and Logit Lens, often produce poor spatial explanations due to weak text-to-vision attention and sink tokens. We propose Angular Relevance (

R_L

), a simple visualization measure based on cosine similarity between visual token embeddings and unembedding-matrix rows. Because

R_L

is invariant to token norm, it produces substantially sharper and more faithful heatmaps than attention maps or Logit Lens on out-of-the-box models. Without any training,

R_L

yields competitive referring-expression segmentation (RES), outperforming prior diagnostic maps and CLIP-based training-free methods. To preserve localized object information during cross-modal mixing, we introduce Angular Object-Aware Loss (AOL), an auxiliary loss that directly supervises visual token embeddings to align angularly with the vocabulary directions of their corresponding object concepts. Finetuning with AOL alongside the standard next-token prediction objective significantly sharpens object localization in heatmaps, showing that the VLM's own visual tokens can be used directly for grounding and improving RES, without training any segmentation heads and without degrading general VQA performance. We further demonstrate that the learned grounding also transfers to category-level segmentation of object categories not seen during training.

Abstract

VLM Interpretability

Explore similar work

Angular Supervision for Visual Object Grounding in Autoregressive Vision-Language Models

Language Models Can Explain Visual Features via Steering

Attention-Guided Saliency Maps for Interpreting Visualization Literacy in VLMs