Spatial-Matrix Attention in Atypical Cognition: Full-Grid Dynamic Scanning and Mid-Sweep Directional Switching as a Model for Next-Generation Non-Linear AI Architectures

Standard large language model (LLM) architectures process natural language sequentially using one-dimensional positional encodings and autoregressive causal masks. While effective for localized sequence generation, this serial pipeline creates severe processing bottlenecks during global structural synthesis, cross-modal translation, and parallel error detection. In contrast, neurodivergent human cognitive processing—specifically visual-spatial dyslexic architectures—treats textual and symbolic input as a two-dimensional spatial coordinate plane. Rather than relying on rigid, single-direction visual paths, this cognitive model deploys simultaneous multi-directional raster sweeps: horizontal (left-to-right, right-to-left), vertical (top-to-bottom, bottom-to-top), diagonal X-pattern scans, and dynamic mid-pass directional pivots upon semantic or structural convergence. This monograph formalizes Spatial-Matrix Attention (SMA), an artificial intelligence attention paradigm that replaces one-dimensional token queues with a 2D Euclidean coordinate tensor. By assigning spatial coordinates to tokens and enabling dynamic vector scanning with early-stopping pointer retrieval, SMA allows machine learning models to execute asynchronous parallel drafting, real-time self-proofreading, and low-latency ontological memory access without requiring heavy autoregressive re-evaluation passes. Main Highlights & Key Contributions: 2D Coordinate Embeddings: Replaces standard 1D sequential indexing with a dynamic 2D Euclidean spatial distance matrix and relative angular orientation tensors. Multi-Vector Matrix Rastering: Formalizes vertical column sweeps, cross-planar diagonal scans, and horizontal parsing for global layout and structural density verification. Mid-Sweep Directional Pivoting: Introduces stochastic angle switching and early stopping during raster passes when key semantic thresholds or structural anomalies are detected. Key-Bound Pointer Retrieval: Enables direct execution of pre-cached ontological macros, drastically reducing KV-cache expansion overhead during long-context reasoning.