The review examines state-space models as possible alternatives to the attention-based designs used in many modern AI systems.
Transformers have performed strongly in large pre-trained AI systems and many other tasks, but their computational demands have encouraged researchers to seek more efficient designs. This survey examines state-space models, which the authors describe as possible replacements for the self-attention component of Transformers.
The paper explains how state-space models work, reviews existing systems and examines their use in natural-language processing, computer vision, graphs, multimodal and multimedia tasks, point clouds and event streams. It also presents statistical comparisons intended to show how different model structures relate to performance across tasks, and identifies directions for further theoretical and applied research.
What the survey covers
The researchers reviewed state-space model studies and applications across several areas of artificial intelligence. Their analysis covers language, images, graphs, multimodal and multimedia data, point clouds and event streams.
The survey also compares different state-space model structures and their reported performance across tasks. It describes these models as a possible way to reduce the complexity associated with attention-based systems, but the abstract does not report a single result showing that state-space models consistently outperform Transformers or use less computing in every setting.
Evidence and limits
This is a survey article with experimental comparisons and analyses of existing state-space models, rather than a report of one new model or one controlled experiment. The abstract does not provide the comparison datasets, numerical results, costs or task-by-task outcomes.
Its conclusions therefore summarize a broad research area and suggest possible advantages and research directions. They do not establish that state-space models are universally cheaper or more accurate than Transformers, or that they will replace them across applications.
// Source
Visual Intelligence · 2026 · DOI: 10.1007/s44267-026-00133-1
Authors: Xiao Wang, Shiao Wang, Yuehang Li, Wentao Wu, Rong Yao, Weizhe Kong, Ju Huang, Shihao Li, Haoxiang Yang, Ziwen Wang, Chenglong Li, Yaowei Wang, Yonghong Tian, Jin Tang
Institutions: Peking University, Harbin Institute of Technology, Anhui University, Peng Cheng Laboratory