An HDC Transformer Architecture for Coherent Narrative Generation

Abstract—Transformers dominate modern sequence modelingbut rely on quadratic-time self-attention and large parame-ter counts, limiting their deployment on resource-constrainedand sovereign systems. This paper introduces an alternativeformulation of the Transformer grounded in HyperdimensionalComputing (HDC) rather than matrix-based attention. Usinghigh-dimensional vector binding, superposition, similarity, andsymbolic constraints, we construct a narrative generation systemcapable of producing coherent stories from raw text. We formallydemonstrate that attention, positional encoding, and residualstreams emerge naturally from high-dimensional geometry, yield-ing a Transformer-equivalent model with linear-time complexity.Empirical results show long-range coherence, entity persistence,and stability under 4-bit MP4C quantization, enabling efficient,interpretable edge AI.Index Terms—Hyperdimensional Computing, TransformerEquivalence, Neuro-Symbolic AI, Edge AI, MP4C Quantization