Proposed SPEAR, a plug-in module inserted between the patch embedding and encoder of a frozen pre-trained ViT. Inspired by human visual attention — where a global glimpse precedes detailed examination — SPEAR pools patch representations into a context vector, scores each patch's importance via a lightweight MLP, then reorders patches using Straight-Through Estimation (STE) for end-to-end differentiable training with a standard cross-entropy loss.

Unlike prior patch-reordering methods that rely on reinforcement learning (e.g., REINFORCE), SPEAR requires no additional loss functions or RL machinery, and trains with frozen backbone weights — updating only its own scorer parameters.

Results: +1.34pp Top-1 on ImageNet-1K and +1.93pp on CIFAR-10 over frozen ViT-Base baselines, demonstrating that complete ViT architectures are order-sensitive in practice despite attention's theoretical permutation-invariance.

Behind the name. The name draws inspiration from the Sphere in Las Vegas, where this project was developed during a research stay at UNLV.