AbstractThe deployment of Large Language Models (LLMs) in browser-based environments faces a fundamental dichotomy: the need for user-specific adaptation versus the static nature of pre-trained weights. While quantization allows models like TinyLlama or Phi-2 to execute on client devices, they remain "read-only" artifacts, unable to learn from immediate context due to the prohibitive cost of backpropagation in WebGPU shaders. This paper proposes a paradigm shift from algebraic matrix retrieval to geometric state evolution. We introduce the Native Object Vector Architecture (N.O.V.A.), a design exploratory that projects context onto a complex spinor manifold using a "Rotation-Gating-Injection" mechanism. Drawing inspiration from Holographic Reduced Representations (HRR), N.O.V.A. encodes semantic relationships as phase differences, enabling an Online Active Inference mechanism via fast Hebbian learning. Experiments on a legacy Intel-based MacBook (2017) validate the architectural efficiency: the model demonstrates a 66× throughput increase compared to a quantized GPT-2 baseline (~47ms vs. ~3131ms per token) while maintaining a minimal 350MB memory footprint. Crucially, empirical analysis reveals the emergence of stable "Persona Vectors"—consistent behavioral modes that persist across generation steps where Transformer baselines exhibit catastrophic collapse. Code and Demos: Github
Yihu Wu (Sun,) studied this question.