Tsinghua and Alibaba Joint Paper Introduces ViT³: A Vision Transformer with Linear Complexity — CVPR 2026 Oral

Tsinghua University and Alibaba have introduced ViT³, a novel vision transformer architecture achieving linear computational complexity. This advancement, presented at CVPR 2026, aims to make high-resolution image understanding feasible on edge devices. The model reinterprets the attention mechanism through Test-Time Training, reducing the traditional quadratic complexity associated with increasing image resolution. By constructing a lightweight internal model during inference, ViT³ enables more efficient information processing.