INTR++: INterpretable TRansformer with better Accuracy, Interpretability, and Adaptability

Dipanjyoti Paul; Arpita Chowdhury; Maneet Chatterjee; Daichi Kitaguchi; Tanya Berger-Wolf; Tetsuya Sakurai; Wei-Lun Chao

cs.CVOct 27, 2026

INTR++: INterpretable TRansformer with better Accuracy, Interpretability, and Adaptability

Authors: Dipanjyoti Paul, Arpita Chowdhury, Maneet Chatterjee, Daichi Kitaguchi, Tanya Berger-Wolf, Tetsuya Sakurai, Wei-Lun Chao

Organizations: IIT Kharagpur · IIEST Shibpur · University of Tsukuba · The Ohio State University ·

Abstract

Interpretable Transformers often perform well in domain specific settings but lose effectiveness beyond that. Achieving ViT-like broad applicability requires models with consistent performance across domains. Recent works on Interpretable Transformers, INTR and Prompt-CAM, were developed keeping domain adaptability in mind. Prompt-CAM leverages learnable prompts to promote adaptability. However, these prompts are fixed in nature and therefore struggle to adapt to domains with substantial variation in input image characteristics, resulting in reduced interpretability. To address this limitation, we propose INTR++, which extends Prompt-CAM by incorporating a lightweight Prompt Adapter Module (PAM). PAM enables prompts to be adaptive by enriching them with domain-specific knowledge, thereby ensuring consistent interpretability across diverse domains including those with varying image characteristics such as medical imaging, micro-scopic imaging, and remote sensing imagery. INTR++ retains the standard architecture and end-to-end learning of Interpretable Transformer based models, while adding only a small number of additional parameters compared to Prompt-CAM. We demonstrate that INTR++ delivers consistent improvements in all aspects across domains like animals, plants, medical imaging, microscopic imaging, remote sensing imagery and object-centric data.

Figures & tables

Figure 2: Architecture of INTR++. INTR++ employs a pretrained frozen backbone, while learning only the prompts and a lightweight prompt adapter module. The module is simple, adding only a small number of additional parameters. It consists of a lower-dimensional linear projection, a non-linear activation, self-attention over prompt tokens, cross-attention with image patches, and a projection back to the original dimension. The adapted prompt tokens at the last layer are finally used for classification. See the Method section for details.

Figure 3: Information induced by different adapters. The first $200$ tokens in the plot are learnable prompt tokens for CUB dataset, while the remaining tokens are image patches. The bold curves represent the exponential moving average (EMA) of the raw cosine similarity values computed using a span of fifteen.

Figure 4: Learned gate values over training epochs. The plot shows the last layer gate values of INTR++ on CUB dataset.

Method	CUB		GBCU
Method	Insertion $\uparrow$	Deletion $\downarrow$	Insertion $\uparrow$	Deletion $\downarrow$
Grad-CAM	0.52	0.17	0.57	0.26
Layer-CAM	0.54	0.13	0.44	0.37
Eigen-CAM	0.56	0.22	0.45	0.39
Attention Rollout	0.55	0.27	0.51	0.26
Prompt-CAM	0.61	0.09	0.59	0.21

Table 1: Quantitative evaluation of interpretation. Lower deletion and higher insertion scores indicate better performance.

Method	CUB	Dog	MedLeaf	Flower	GBCU	Brain	Colon	Car	Craft	AID
Prompt-CAM	73.2	81.1	96.9	86.8	63.1	93.5	74.2	83.2	68.7	93.5
INTR++	79.0	83.5	97.1	87.2	81.2	99.2	84.4	88.1	75.4	97.7

Table 2: Accuracy comparison. Results are on the DINO backbone. INTR++ outperforms Prompt-CAM across domains.

Figure 5: Qualitative interpretability. Comparison between INTR++ and other post-hoc methods on the GBCU dataset.

Figure 6: Trait manipulation. Top row: the white-neck trait highlighted as important for the White-necked Raven. Bottom row: after deleting this trait, the model predicts American Crow and no longer considers the white neck as important.

Figure 7: Comparison of interpretable models. Visual demonstration (heatmaps) of the four strongest responses on a “Groove Billed Ani” example image.

Figure 8: INTR++ and Prompt-CAM interpretation comparisons across all domains. Red markings denote regions of interest.

		CUB	Car	Flower	GBCU	Colon	AID
Ours	DINO	79.0	88.1	87.2	81.2	84.4	97.7
	DINOv2	79.6	91.3	87.4	82.0	84.7	96.6
	BioCLIP-2	91.1	94.2	96.6	82.3	84.9	97.4

Table 3: Accuracy of INTR++ on different backbones. We consider one dataset from each of the animal, object, plant and medical domain to demonstrate its effectiveness across backbones.

Figure 9: INTR++ on different backbones. INTR++ can identify important traits for differentiating Heermann Gull species irrespective of the backbone used.

	CUB	Car	Flower	GBCU	Colon	AID
INTR-MLP	74.0	67.2	62.6	51.7	72.4	89.0
INTR-AdaptFormer	72.6	81.0	76.7	41.0	80.9	96.8
INTR-LoRA	75.0	83.0	85.1	55.7	81.2	97.3
INTR++	79.0	88.1	87.2	81.2	84.4	97.7

Table 4: Accuracy comparison. INTR++ with other adapters.

Figure 10: Comparison with adapters. INTR++ highlights important traits, while others produce noisier outputs.

Method	CUB	Car	Flower	GBCU	Colon	AID
INTR++ $\setminus$ {PCA}	77.3	87.1	85.1	80.8	82.5	97.1
INTR++ $\setminus$ {PSA}	73.3	85.8	86.7	79.2	82.1	96.0
INTR++ $\setminus$ {PSA, PCA}	72.6	82.8	83.4	41.0	74.5	96.9
INTR++	79.0	88.1	87.2	81.2	84.4	97.7

Table 5: Ablation study of INTR++. The study shows the significance of each of the components of the Prompt Adapter.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 11: Design of adapter AdaptFormer. The module applies a lower-dimensional projection, a non-linear activation, and a projection back to the original dimension.

Dataset	CUB	Dog	MedLeaf	Flower	Car	Craft
# Train Images	5,994	12,000	1,455	2,040	8144	3334
# Test Images	5,795	8,580	380	6,149	8041	3333
# Classes	200	120	30	102	196	100

Appendix

Table 6: Characteristics of the natural image datasets for the Animal, Plant, and Object-centric domains.

Dataset	GBCU	Brain	Colon	AID
# Train Images	1133	5712	4928	8000
# Test Images	122	1311	1232	2000
# Classes	3	4	8	30

Appendix

Table 7: Characteristics of the specialized image datasets for the Medical, Microscopic, and Remote-Sensing Imagery domains.

Figure 12: Comparison of INTR++ with post-hoc methods. INTR++ shows the top two attention maps for correct prediction.

Figure 13: Interpretation results of INTR++ across different domains. We show the top four attention maps for each input image. It highlights the model’s reliability in focusing on important traits or attributes for prediction across natural and specialized image domains.

Figure 14: Interpretation comparisons on specialized images. We compare INTR++ with other interpretable methods on the GBCU dataset for the benign category.

Figure 15: INTR++ predicts correctly where Prompt-CAM fails. We visualize all twelve attention heads for both approaches. INTR++ effectively captures the key traits and yields accurate predictions, whereas Prompt-CAM predominantly attends to redundant traits across multiple heads and includes noisy regions, resulting in incorrect predictions.

Figure 16: Interpretation of INTR++ and Prompt-CAM . We show comparisons on the MedLeaf, Dog and Craft datasets.

Figure 17: Attribute manipulation. The top row shows the important cancerous region for a malignant case; deleting it leads to a normal prediction, with attention shifting elsewhere.

	CUB	Car	Flower	GBCU	Colon	AID
ProtoConcepts	68.3	78.7	85.5	41.8	78.9	95.1
PIPNet	78.8	84.4	88.2	79.9	81.8	95.4
TesNet	58.9	58.5	50.3	77.8	80.2	94.9
INTR++	79.0	88.1	87.2	81.2	84.4	97.7

Appendix

Table 8: Accuracy comparisons. INTR++ with other interpretable models. We consider one dataset from each domain.

Method	CUB		Flower		Car
Method	Insert. $\uparrow$	Delet. $\downarrow$	Insert. $\uparrow$	Delet. $\downarrow$	Insert. $\uparrow$	Delet. $\downarrow$
ProtoConcepts	0.11	0.76	0.36	0.61	0.48	0.15
PIPNet	0.59	0.10	0.76	0.29	0.79	0.12
TesNet	0.43	0.18	0.24	0.16	0.50	0.09
INTR++	0.62	0.08	0.66	0.18	0.82	0.04

Appendix

Table 9: Evaluation of interpretation for natural image datasets. Comparisons among interpretable methods

Method	GBCU		Colon		AID
Method	Insert. $\uparrow$	Delet. $\downarrow$	Insert. $\uparrow$	Delet. $\downarrow$	Insert. $\uparrow$	Delet. $\downarrow$
ProtoConcepts	0.52	0.45	0.45	0.55	0.50	0.66
PIPNet	0.63	0.52	0.75	0.59	0.66	0.64
TesNet	0.58	0.55	0.75	0.43	0.65	0.66
INTR++	0.65	0.19	0.77	0.39	0.66	0.50

Appendix

Table 10: Evaluation of interpretation for specialized image datasets. Comparisons among interpretable methods

Figure 18: Ablation study on GBCU dataset. Visual comparison demonstrating the importance of each component in the PAM of INTR++ on the specialized images input.

Figure 19: Ablation study on CUB dataset. Visual comparison demonstrating the importance of each component in the PAM of INTR++ on the natural images input.

Figure 20: Learned gate values over training epochs. The plot shows the last layer gate values of INTR++ on GBCU dataset.

Figure 21: Projection size comparison. Performance comparison with different projection sizes for INTR++, INTR-LoRA, and INTR-AdaptFormer for the CUB dataset.

Figure 22: Interpretation of MLP-based adapters. Top attributes for an image from the CUB and GBCU datasets.

Dataset	INTR++	Prompt-CAM	$p$ -value
CUB	78.2 $\pm$ 1.1	74.3 $\pm$ 2.21	0.011
GBCU	80.4 $\pm$ 1.2	62.8 $\pm$ 1.71	0.002

Appendix

Table 11: Statistical analysis. Mean

\pm

standard deviation over ten runs.

p

-values are from the paired Wilcoxon signed-rank test between INTR++ and Prompt-CAM on CUB and GBCU datasets.

ProtoConcepts	PIPNet	TesNet	Prompt-CAM	INTR++
$\approx 85$ – $86$	$\approx 86$ – $87$	$\approx 86$ – $87$	$\approx 0.1$ – $1.0$	$\approx 1.2$ – $3.4$

Appendix

Table 12: Number of learnable parameters (in millions). Comparison of interpretable methods in terms of parameter efficiency using DINO as the backbone. Values indicate the approximate range across all datasets.

Transformer Interpretability

Transformer interpretability studies how attention, feed-forward layers, and hidden representations support model predictions and behavior.

See all 340 papers

From Adaptation to Generalization: Adaptive Visual Prompting for Medical Image Segmentation
Apr 19, 2026Evren Çetinkaya, Sangmin Lee, Jung Uk Kim +2Domain Generalization Medical Image Segmentation
Technical University of Munich · Korea University · Kyung Hee University +1
UniPrompt-CL: Efficient Single-Inference Continual Learning for Medical Image Classification
Aug 14, 2025Gyutae Oh, Jitae ShinContinual Learning Domain-Incremental Learning
Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon 16419, Republic of Korea
Joint Semantic Token Selection and Prompt Optimization for Interpretable Prompt Learning
May 6, 2026Yating Wang, Yaqi Zhao, Yongshun Gong +2Prompt Learning VLM Adaptation
School of Software, Shandong University, Jinan, 250101, China

Apr 19, 2026J/K move · Enter open · S save

From Adaptation to Generalization: Adaptive Visual Prompting for Medical Image Segmentation

Evren Çetinkaya, Sangmin Lee, Jung Uk Kim, Hong Joo Lee+1

Technical University of Munich · Korea University · Kyung Hee University+1

Domain Generalization Medical Image Segmentation Visual Prompt Learning

Open paper AskPDF arXiv

Visual prompting has emerged as a powerful method for adapting pre-trained models to new domains without updating model parameters. However, existing prompting methods typically optimize a single prompt per domain and apply it uniformly to all inputs, limiting their ability to generalize under intra and inter-domain variability, which is especially critical in the medical field. To address this, we propose APEX, an Adaptive Prompt EXtraction framework that retrieves input-specific prompts from a learnable prompt memory. The memory stores diverse, domain-discriminative prompt representations and is queried via domain features extracted from the Fourier spectrum. To learn robust and discriminative domain features, we introduce a novel Low-Frequency Feature Contrastive (LFC) learning framework that clusters representations from the same domain while separating those from different domains. Extensive experiments on two medical segmentation tasks demonstrate that APEX significantly improves generalization across both seen and unseen domains. Furthermore, it complements any existing backbones and consistently enhances performance, confirming its effectiveness as a plug-and-play prompting solution in medical fields. The code is available at https://github.com/cetinkayaevren/apex/

Abstract

Transformer Interpretability

Explore similar work

From Adaptation to Generalization: Adaptive Visual Prompting for Medical Image Segmentation

UniPrompt-CL: Efficient Single-Inference Continual Learning for Medical Image Classification

Joint Semantic Token Selection and Prompt Optimization for Interpretable Prompt Learning