The paper presents a detailed study of feature representations in Vision Transformers (ViTs), revealing that feature collapse at initialization causes redundancy and proposing a reduction scheme to address it. It demonstrates that entropy and minimum eigenvalue metrics reliably predict generalization, and that token‑space features and linear submodule outputs are particularly informative. The authors develop a proxy that improves ranking of ViT architectures by 18‑48% over previous baselines, enabling the selection of models with higher accuracy at comparable or lower computational cost.
arXiv:2606. 08156v1 Announce Type: cross Abstract: Vision Transformers (ViTs) achieve strong performance but suffer from high computational costs due to quadratic self-attention complexity.
arXiv:2608. 20134v1 Announce Type: cross Abstract: We present a novel view on feature evolution in Vision Transformers (ViTs) by visualizing the training process over two dimensions -- network depth (layer) and training time (epochs).
The paper studies Graph-Guided Token Merging (G2TM), a module that reduces token count in Vision Transformers. It evaluates G2TM across multiple segmentation frameworks and decoder types, finding that its performance gains are tied to the encoder rather than the decoder. The authors report consistent reductions in GFLOPs (22‑47%) and throughput improvements (up to 74%) on ADE20K, with optimal hyperparameters depending mainly on backbone pre‑training and target dataset.
The paper introduces Prune-and-Merge ViT, a training‑free token compression framework for Vision Transformers that combines class‑attention pruning with similarity‑based merging to reduce computational cost while preserving semantic information. It evaluates robustness using two new metrics—Relative Accuracy Drop (RAD) and Prediction Flip Rate (PFR)—and shows that the method maintains or improves accuracy on pretrained ViT and DeiT models compared to existing token reduction techniques. The authors also argue that the hybrid approach does not induce hallucination‑like behavior and keeps predictions stable under common image transformations.
HiLRP introduces a unified attribution framework for Vision Transformers (ViTs) that addresses the challenges posed by diverse architectural designs. By decomposing ViT operations into four basic types—linear maps, bilinear mixing, normalization/gating, and reindexing—HiLRP applies conservation‑satisfying relevance rules, enabling reliable explanations across a wide range of backbones. The method outperforms 14 existing attribution techniques on 10 architectures, maintaining conservation and improving localization accuracy (0.97 Pointing) compared to competitors.
arXiv:2607. 10391v1 Announce Type: cross Abstract: Despite exposing rich intermediate representations, Vision Transformers (ViTs) are almost exclusively utilized as black-box feature extractors, where only the last layer is considered for downstream tasks.
arXiv:2605.18177v2 Announce Type: replace Abstract: Query-based Vision Transformer segmentation models typically reconstruct dense spatial feature maps to predict masks, inheriting design patterns fr...
arXiv:2606. 01503v1 Announce Type: cross Abstract: Unified vision-language models (VLMs) integrate visual understanding and visual generation within a single autoregressive backbone, but their joint training is computationally expensive and largely overlooked from an efficiency perspective.
arXiv:2607. 02386v1 Announce Type: cross Abstract: While Vision Transformers have achieved remarkable success across computer vision and language applications, the geometric evolution of their internal representations throughout training remains insufficiently understood.
arXiv:2608. 06901v1 Announce Type: cross Abstract: Vision-language models (VLMs) have achieved remarkable generalization across diverse multimodal tasks through large-scale pre-training, yet their rapidly increasing computational and memory requirements pose significant challenges for deployment in constrained environments.
The paper revisits feature inversion via inverse networks, proposing a modular variant that is computationally efficient and can produce more semantically coherent image reconstructions. It applies this method to large‑scale transformer‑based vision models—Detection Transformer, Vision Transformer, Swin Transformer, and Data‑Efficient Image Transformer—examining reconstructions across network depth. The analysis highlights gradual representational changes in Detection and Vision Transformers, systematic differences in preserving contextual shape and fine‑grained details, and varying robustness to color perturbations, while also demonstrating the framework’s broader applicability to other transformer architectures.
arXiv:2512. 08240v2 Announce Type: replace-cross Abstract: Vision-language models (VLMs) rely on hundreds of visual tokens, leading to high computational and memory costs.