E8 Lattice Quantization for State Space Model Hidden States
Abstract
E8 lattice quantization applied to Mamba SSM hidden states. At 4-bit, E8 preserves 99% of baseline accuracy (29.89% vs 30.18%) where scalar quantization produces 0.00%. At 2-bit, E8 retains 87% of baseline (26.24%). First application of lattice quantization to SSM hidden states. No model retraining or architectural changes required. Code: https://github.com/Dawizzer/e8-ssm-quantization
Fields of Study & Subjects
Advanced Neural Network Applications
Advanced Data Compression Techniques
Topic Modeling
Publication details
- Publication date
- 12 March 2026
- Language
- EN
- Document type
- Preprint