Research Papers¶
This section documents the theoretical foundations and key innovations behind Bit-Axon. Each page provides the mathematical formulation, design rationale, and implementation mapping for a core component of the system.
Papers¶
| # | Paper | Status | Key Idea |
|---|---|---|---|
| 1 | Axon-SSM: Selective SSM for Apple Silicon |  | Mamba-style selective state space model with @mx.compile fused kernels, \(\mathcal{O}(1)\) memory per token |
| 2 | 24-Layer Sandwich Architecture | | Three-zone hybrid: SSM → SWA+MoE → SSM+MoE, dimension bridge \(d_{\text{src}}=2048\) |
| 3 | Thermal-Aware Training | | CoolingScheduler + macOS powermetrics, pause at 85°C, halt at 95°C |
| 4 | TurboQuant KV Cache Compression |  | Compress KV cache for 64K contexts; ICLR 2026 reference |
Scope¶
These papers focus on the mathematical foundations and algorithmic design of each component. For API usage and integration details, see the Architecture section.
Referenced Work¶
- Mamba: Gu, A., & Dao, T. (2023). Mamba: Linear-Time Sequence Modeling with Selective State Spaces. arXiv:2312.00752.
- MLX: Apple Machine Learning Research. MLX: An array framework for machine learning on Apple silicon.
- Qwen2.5: Qwen Team (2024). Qwen2.5 Technical Report.
See also¶
- Architecture Overview — Implementation details of each component
- Training Guide — Thermal-aware QLoRA training pipeline
- Quantization Guide — NF4 quantization and merge workflows