A Smaller Transformer in Your Transformer

Vision Transformers (ViTs) are built by stacking independently parameterized blocks, but it remains unclear how much of this depth requires layer specific transformations and how much can be realized through recurrent computation. We study this question with bViT, a single-block recurrent ViT in which one transformer block is applied repeatedly to process an image. This architecture preserves the iterative structure of a deep ViT while removing layer specific block parameterization, providing a controlled setting for studying recurrence in vision. On ImageNet-1K, a 12-step bViT-B achieves accuracy comparable to standard ViT-B under the same training recipe and computational budget, while using an order of magnitude fewer parameters. We observe that recurrent performance improves with representation width, with wider bViTs recovering much more of the performance of standard ViTs than narrow variants. We interpret this behavior as implicit depth multiplexing, where a shared block expresses multiple step-dependent computations through the evolving hidden state. Beyond ImageNet classification, bViT transfers competitively to downstream tasks and enables parameter-efficient fine-tuning. Mechanistic analyses of activations, attention and step-specific pruning show that the shared block changes its effective behavior across recurrent steps rather than simply repeating the same computation. Our results suggest that a large fraction of ViT depth can be implemented through recurrent reuse, provided that the representation space is sufficiently wide.

While prior studies have successfully compressed vision Transformers (ViTs) through various pruning techniques, most have concentrated on width pruning to achieve significant reductions in model size. Depth pruning, which removes entire layers from a ViT, is notoriously difficult for accuracy recovery despite its potential to deliver higher speedups, limiting the acceleration achieved by existing joint width-and-depth pruning methods. In this work, we reveal that the failure of existing depth pruning methods lies in their neglect of heterogeneity between different layers, and we introduce HetDPT, a heterogeneity-aware depth pruning method that avoids dimension mismatch. Comprehensive experiments on ImageNet-1K, CIFAR-100, COCO, and ADE20K validate our method: HetDPT achieves a 1.58$\times$ speedup for DeiT-B while maintaining accuracy and a 1.39$\times$ speedup for DeiT-S with nearly no accuracy degradation. Furthermore, when combined with width pruning, HetDPT+ sets a new state-of-the-art record in extreme ViT pruning, enhancing the acceleration ratio from 4.24$\times$ to 5.19$\times$ for the Isomorphic-Pruning-2.6G configuration while maintaining near-lossless accuracy; our code is available at https://github.com/Efficient-AI-for-All/HetDPT.

Vision Transformers (ViTs) achieve state-of-the-art segmentation accuracy but require large training datasets because each layer has unique parameters that must be learned independently. We present RD-ViT, a Recurrent-Depth Vision Transformer that adapts the Recurrent-Depth Transformer (RDT) architecture to dense prediction tasks, supporting both 2D and 3D inputs. RD-ViT replaces the deep stack of unique transformer blocks with a single shared block looped T times, augmented with LTI-stable state injection for guaranteed convergence, Adaptive Computation Time (ACT) for spatial compute allocation, depth-wise LoRA adaptation, and optional Mixture-of-Experts (MoE) feed-forward networks for category-specific specialization. We evaluate on the ACDC cardiac MRI segmentation benchmark in both 2D slice-level and 3D volumetric settings with exclusively real experiments executed in Google Colab. In 2D, RD-ViT outperforms standard ViT at 10% training data (Dice 0.774 vs 0.762) and at full data (0.882 vs 0.872). In 3D, RD-ViT with MoE achieves Dice 0.812 with 3.0M parameters, reaching 99.4% of standard ViT performance (0.817) at 53% of the parameter count. MoE expert utilization analysis reveals that different experts spontaneously specialize for different cardiac structures (RV, MYO, LV) without explicit routing supervision. ACT halting maps show higher compute allocation at cardiac boundaries, and the mean ponder time decreases from 2.6 to 1.4 iterations during training, demonstrating learned computational efficiency. Depth extrapolation enables inference with more loops than training without degradation. All code, notebooks, and results are publicly released.