Abstract
Vision Transformers have become the dominant paradigm for visual learning due to their ability to model long-range dependencies through self-attention. However, dense pairwise attention introduces quadratic computational complexity with respect to token count, severely limiting scalability in high-resolution and long-duration perception tasks. Existing efficiency improvements reduce attention overhead but preserve the underlying assumption that global visual understanding requires explicit token interaction.
This paper introduces the SpectraGrid Vision Transformer (SG-ViT), a unified architecture that replaces dense self-attention with structured perceptual computation, sparse routing, persistent latent world modeling, and autonomous action-driven reasoning. Instead of recomputing dense token interactions at every timestep, SG-ViT maintains a persistent structured latent representation of the environment and updates it incrementally through sparse observation-driven inference.
The architecture integrates three core principles: structured spatial decomposition for efficient local perception, persistent latent world memory for temporal reasoning, and closed-loop planning mechanisms for autonomous interaction with dynamic environments. This formulation transforms perception from sequence processing into structured world-state evolution, enabling long-horizon reasoning, object permanence, predictive simulation, and autonomous planning while achieving near-linear scaling in both spatial and temporal dimensions.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Manuel-Devadoss, Johny and Prasad, Deepthi, "SpectraGrid Vision Transformer", Technical Disclosure Commons, (May 11, 2026)
https://www.tdcommons.org/dpubs_series/10058