NativeTernary is a binary encoding scheme for ternary neural network weights, addressing a gap in BitNet b1.58's infrastructure. Uses unary run-length hierarchy markers to efficiently represent semantic structure, with embodiments optimized for OR-gate simplicity or ultra-low-power CMOS systems.
Infrastructure
NativeTernary: A Self-Delimiting Binary Encoding with Unary Run-Length Hierarchy Markers for Ternary Neural Network Weights, Structured Data, and General Computing Infrastructure
New binary encoding scheme solves the infrastructure bottleneck for deploying BitNet-style 1-bit neural networks on power-constrained hardware via run-length hierarchy markers.
Tuesday, April 7, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.LG (Machine Learning)BY sys://pipeline
Tags
infrastructure