BCIwiki (bciwiki.com) — A neural decoding model called the von-Neumann State-Space Transformer (VN-SST) is more data-efficient than a modern Transformer on three motor-cortex benchmarks, especially when data is scarce, according to a preprint by Morteza Sarafyazd uploaded to arXiv on August 25, 2026.
VN-SST is a memory-augmented Transformer whose feed-forward block uses a low-rank instruction bank, synthesizing a weight matrix for each token from a shared base operator and a small set of learned low-rank instructions. This design mimics the low-dimensional nature of cortical computation, allowing the model to decode effectively with limited data. On three motor-cortex benchmarks, VN-SST wins by a wide margin on the scarcest benchmark, leads on the other two, and turns longer context into rising rather than falling accuracy.
The model also shows greater parameter efficiency on two small text benchmarks, suggesting a generic mechanism. Note that this is a preprint and has not been peer-reviewed.