A Systematic Review of Swin Transformer, Mamba, and Kolmogorov–Arnold Network Architectures and Their Hybrids in Visual Computing: Retrospective Registration

Abstract

Retrospective registration. This protocol is registered after the review was completed: the literature search was last run on 31 March 2026, and screening, data extraction and synthesis had been finished before this registration was submitted. It records the review question, eligibility criteria, sources, search strategy, selection, data extraction and synthesis methods as they were applied, so that the conduct of the review can be checked against the published report. Purpose. Since 2021, three architecture families have each changed one building block of the Transformer: the Swin Transformer confines self-attention to shifted local windows, Mamba and other selective state-space models (SSMs) replace attention with an input-dependent recurrence, and Kolmogorov–Arnold Networks (KANs) move the nonlinearity of the multilayer perceptron to learnable functions on the edges. No previous review covers the three families jointly, and none quantifies how consistently primary studies report the computational efficiency they claim. This systematic review compares the three families and their hybrids in visual computing, including biomedical imaging, industrial and remote-sensing imagery, LiDAR and point-cloud data. Research questions. RQ1. What does each family contribute, in terms of what it represents, where it innovates and why it improves on the block it replaces? RQ2. How do the families differ in representational power and computational efficiency? RQ3. How consistently does the literature support the efficiency it claims? RQ4. How far has the combination of the families been taken? Methods (summary). Six sources (arXiv, Crossref, OpenAlex, PubMed, Semantic Scholar and DBLP) were searched through their public programmatic interfaces for English-language studies published from 2021 to 2026. Records were screened by title and abstract, prioritised with a score combining venue quality, citations and level of contribution, and the highest-ranked records were assessed in full text. Included studies were extracted with a common schema and compared along six axes: architectural variants, applications, computational efficiency, interpretability, data efficiency and hybridisation. Synthesis is narrative and tabular; no meta-analysis is performed. The review is reported in accordance with PRISMA 2020. Expected outcomes. A comparative map of the contribution of each family and their hybrids, the tasks and domains where each is used, and an assessment of how often studies report parameters, FLOPs, latency, memory and code, leading to recommendations for standardised efficiency reporting.