The paper introduces reViT, a recurrent vision transformer that uses a single Transformer block repeatedly to achieve the accuracy of a full-depth encoder while keeping inference FLOPs comparable. It achieves depth‑specific transformations by expressing each block’s feed‑forward network as a convex combination of a small shared expert bank, with a continuous depth coordinate guiding the mixture. Experiments on ImageNet‑1k and DINOv2 distillation show that reViT matches or surpasses state‑of‑the‑art models, reduces stored parameters by about 70%, and supports elastic‑depth deployment from a single checkpoint.
arXiv:2608. 15062v1 Announce Type: cross Abstract: Scaling transformer language models creates an inherent tension between expressivity and memory efficiency.
arXiv:2608.15062v3 Announce Type: replace-cross Abstract: Scaling transformer language models creates an inherent tension between expressivity and memory efficiency. While unique weights across layer...
The paper introduces Gated Recurrent Transformers, a depth‑sharing architecture that brackets a single shared core with fixed prelude and coda blocks and uses a lightweight projection and element‑wise update gate to modulate recurrent updates. This design allows functional specialization across recurrences while reducing memory footprint. Experiments show that, under equal FLOPs or parameter budgets, the recurrent model matches or surpasses deeper GPT‑2 Small baselines, achieving similar or better accuracy with fewer parameters and lower peak decoding memory.
The paper introduces Gated Token Recurrence (GTR), a softmax‑free recurrent vision backbone that replaces global softmax attention with gated linear attention, alternating scan directions, and enhanced SwiGLU blocks. GTR is distilled from a DINOv3 teacher using only final‑layer patch‑token alignment, and achieves strong performance on COCO object detection (58.9 box AP) with very low latency (1.908 ms on an RTX 4090). The backbone also transfers to multiple dense prediction tasks and runs efficiently on edge hardware via a specialized CUDA operator and TensorRT deployment.
arXiv:2606. 16112v1 Announce Type: cross Abstract: Residual architectures are ubiquitous in deep learning, but they suffer from a subtle structural limitation: the norm of the residual stream can grow rapidly with depth.
The paper introduces DIFTA-3D, a method that replaces the task‑specific visual branch in IIFNet3D with a frozen DINOv3 foundation model for RGB‑D 3D instance detection. It employs a depth‑consistent feature pipeline that projects points into calibrated RGB‑D frames, filters features with a metric depth‑residual check, caches accepted DINOv3 features, and aggregates them within proposal‑aligned RoI grids. Extensive experiments on ScanNetV2 show that the DINOv3 control achieves mAP scores of 76.15/60.93 at IoU thresholds 0.25/0.50, while the Conservative VAID recipe improves these to 76.59/62.16, indicating a modest gain from the proposed transfer recipe.
MoE-ViE introduces a Mixture-of-Experts vision encoder that scales efficiently for image and video understanding, outperforming dense counterparts across various sizes. The study shows fine‑grained MoE topologies provide significant gains, and proposes an auxiliary‑loss‑free balancing variant and a specialized MoE kernel to reduce inference latency. With frame‑level distillation and a novel freezing mechanism, the largest MoE‑ViE model matches state‑of‑the‑art zero‑shot performance while being 1.7× larger and 76% faster, and it outperforms other encoders when paired with a language model on both image and video benchmarks.
arXiv:2601.18923v2 Announce Type: replace-cross Abstract: Depth sensors are widely deployed across robotic platforms, and advances in fast, high-fidelity depth simulation have enabled robotic policie...
arXiv:2607. 02612v1 Announce Type: cross Abstract: Vision Transformers achieve strong image classification accuracy but process all image regions with nearly the same computation, even when many regions are redundant or uninformative.
ProgResViT is an input‑adaptive Vision Transformer that processes images progressively across multiple rounds, starting with a low‑resolution image and a narrow subnetwork and refining the prediction with higher resolution and a wider subnetwork if needed. The method introduces Progress‑Conditioned Soft Gating (PSG) to share a single backbone across rounds while conditioning token fusion and layer outputs on the current round, block, and input resolution. Experiments on DeiT show improved accuracy‑compute trade‑offs compared to adaptive‑width, adaptive‑depth, and dynamic‑token baselines, and the design also benefits self‑supervised DINO representations and downstream semantic segmentation.
arXiv:2602. 07494v2 Announce Type: replace Abstract: Deeper modern architectures are costly to train, making hyperparameter transfer preferable to expensive repeated tuning.
arXiv:2609.31620v1 Announce Type: new Abstract: Representation autoencoders (RAEs) reuse features from a pretrained visual encoder as reconstruction and diffusion latents, integrating strong visual r...