mlsanityData to 14 Oct 2026

TAB: Transformer Attention Bottlenecks enable User Intervention and Debugging in Vision-Language Models

arXiv 2412.18675 ↗First posted 24 Dec 2024cs.CVcited by 2

On the map: Multimodal LLMs / vision-language models → Visual token pruning and efficiency

Abstract (arXiv)

Multi-head self-attention (MHSA) is a key component of Transformers, a widely popular architecture in both language and vision. Multiple heads intuitively enable different parallel processes over the same input. Yet, they also obscure the attribution of each input patch to the output of a model. We propose a novel 1-head Transformer Attention Bottleneck (TAB) layer, inserted after the traditional MHSA architecture, to serve as an attention bottleneck for interpretability and intervention. Unlike standard self-attention, TAB constrains the total attention over all patches to ∈ [0, 1]. That is, when the total attention is 0, no visual information is propagated further into the network, and the vision-language model (VLM) would default to a generic, image-independent response. To demonstrate the advantages of TAB, we train VLMs with TAB to perform image-difference captioning. Over three datasets, our models perform similarly to baseline VLMs in captioning but the bottleneck is superior in localizing changes and in identifying when no changes occur. TAB is the first architecture to enable users to debug by editing attention, which often produces expected outputs by VLMs.

Key facts. Cited by 2 papers in the map (data through 14 Oct 2026). Placed in Visual token pruning and efficiency, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 52 of its 80 references that are in the map. First posted on arXiv 24 Dec 2024.

Builds on

52 of 80 references.

References · 52

  1. 2017Attention Is All You Needcited by 60,361
  2. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  3. 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 30,186
  4. 2021Emerging Properties in Self-Supervised Vision Transformerscited by 6,968
  5. 2019BERTScore: Evaluating Text Generation with BERTcited by 5,512
  6. 2016SGDR: Stochastic Gradient Descent with Warm Restartscited by 5,045
  7. 2023Improved Baselines with Visual Instruction Tuningcited by 4,418
  8. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  9. 2015Learning Deep Features for Discriminative Localizationcited by 2,436
  10. 2019ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Taskscited by 2,222
  11. 2014CIDEr: Consensus-based Image Description Evaluationcited by 2,046
  12. 2021ViLT: Vision-and-Language Transformer Without Convolution or Region Supervisioncited by 1,230
38 more
  1. 2020Concept Bottleneck Modelscited by 1,032
  2. 2021Scaling Vision Transformerscited by 999
  3. 2023Vision Transformers Need Registerscited by 871
  4. 2021CrossViT: Cross-Attention Multi-Scale Vision Transformer for Image Classificationcited by 806
  5. 2020Quantifying Attention Flow in Transformerscited by 775
  6. 2021Resolution-robust Large Mask Inpainting with Fourier Convolutionscited by 738
  7. 2024Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMscited by 735
  8. 2017Interpretable Explanations of Black Boxes by Meaningful Perturbationcited by 612
  9. 2022Memory-Based Model Editing at Scalecited by 447
  10. 2022Not All Patches are What You Need: Expediting Vision Transformers via Token Reorganizationscited by 363
  11. 2016Top-down Neural Attention by Excitation Backpropcited by 321
  12. 2021Generic Attention-model Explainability for Interpreting Bi-Modal and Encoder-Decoder Transformerscited by 320
  13. 2020Understanding the Role of Individual Units in a Deep Neural Networkcited by 255
  14. 2020Concept Whitening for Interpretable Image Recognitioncited by 225
  15. 2023OpenAGI: When LLM Meets Domain Expertscited by 197
  16. 2020Debugging Tests for Model Explanationscited by 147
  17. 2022ReCLIP: A Strong Zero-Shot Baseline for Referring Expression Comprehensioncited by 120
  18. 2023Dynamic Graph Enhanced Contrastive Learning for Chest X-ray Report Generationcited by 118
  19. 2018Learning to Describe Differences Between Pairs of Similar Imagescited by 114
  20. 2020Rewriting a Deep Generative Modelcited by 107
  21. 2022Better plain ViT baselines for ImageNet-1kcited by 98
  22. 2021The effectiveness of feature attribution methods and its correlation with automatic evaluation scorescited by 88
  23. 2023Which Tokens to Use? Investigating Token Reduction in Vision Transformerscited by 60
  24. 2020Now You See Me (CME): Concept-based Model Extractioncited by 51
  25. 2019Expressing Visual Relationships via Languagecited by 49
  26. 2019Interpretability Beyond Classification Output: Semantic Bottleneck Networkscited by 49
  27. 2023Learning Bottleneck Concepts in Image Classificationcited by 49
  28. 2021Describing and Localizing Multiple Changes with Transformerscited by 47
  29. 2023Expert Knowledge-Aware Image Difference Graph Representation Learning for Difference-Aware Medical Visual Question Answeringcited by 41
  30. 2022Image Difference Captioning with Pre-training and Contrastive Learningcited by 39
  31. 2022CLIP4IDC: CLIP for Image Difference Captioningcited by 25
  32. 2021DeepFace-EMD: Re-ranking Using Patch-wise Earth Mover's Distance Improves Out-Of-Distribution Face Identificationcited by 20
  33. 2024Img-Diff: Contrastive Data Synthesis for Multimodal Large Language Modelscited by 17
  34. 2024OneDiff: A Generalist Model for Image Difference Captioningcited by 15
  35. 2024VIXEN: Visual Text Comparison Network for Image Difference Captioningcited by 12
  36. 2022How explainable are adversarially-robust CNNs?cited by 9
  37. 2021L2C: Describing Visual Differences Needs Semantic Understanding of Individualscited by 8
  38. 2024PEEB: Part-based Image Classifiers with an Explainable and Editable Language Bottleneckcited by 7

50 most cited of 52 All 52 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026Legible-by-Construction: Attention and End-to-End Transformerscited by 3
  2. 2026Imagine How To Change: Explicit Procedure Modeling for Change Captioningcited by 1

Similar, not cited

Close in wording, no citation either way

  1. 2025Head Pursuit: Probing Attention Specialization in Multimodal Transformers
  2. 2025BlindSight: Harnessing Sparsity for Efficient Vision-Language Models
  3. 2025CASA: Cross-Attention over Self-Attention for Efficient Vision-Language Fusion
  4. 2026Vision Token Reduction via Attention-Driven Self-Compression for Efficient Multimodal Large Language Models
  5. 2025See What You Are Told: Visual Attention Sink in Large Multimodal Models
  6. 2024A-VL: Adaptive Attention for Large Vision-Language Models
  7. 2025PEVLM: Parallel Encoding for Vision-Language Models
  8. 2026Attention Sinks in VLMs Are Borrowed, Not Seen