AViGPT: Decoupling Neural Reasoning from Parametric Memory via a Sub-Millisecond Native NVMe Hardware Bus

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Autoregressive language models scale parameter counts to memorize factual data within dense neural connections. Storing encyclopedic associations inside model weights creates substantial memory footprints, high inference costs, and fixed knowledge cutoffs. AViGPT is an autoregressive transformer architecture that decouples syntactic language reasoning from factual storage. The system pairs a 183-million parameter neural core with a native hardware memory bus directly interfacing with local NVMe SSD storage. An explicit nine-token instruction set coordinates execution: the model pauses generation, queries local SQLite FTS5 storage, retrieves verified context in 1.18 milliseconds, and completes responses with verified factual precision. The architecture was pretrained on 5.0 billion tokens of English Wikipedia and FineWeb-Edu text, then aligned on 25,850 task trajectories. Empirical benchmarks confirm a 0.4 GB RAM footprint on consumer CPUs, sub-2ms storage retrieval latency, deterministic arithmetic execution, and immediate zero-shot knowledge insertion via direct disk writes without weight modification.