Reproduce “Scaling Laws for Precision in High-Dimensional Linear Regression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dechen Zhang, Xuan Tang, Yingyu Liang, Difan Zou
This theoretical study at The University of Hong Kong establishes scaling laws for low-precision training in high-dimensional linear regression, revealing that multiplicative quantization (like FP8) preserves model capacity, while additive quantization (like INT8) reduces it. Both schemes decrease effective data size and introduce an additive error floor, providing a mechanistic explanation for different scaling behaviors observed in large language models.
Key insights
- A critical dichotomy exists: multiplicative quantization preserves the full effective model capacity, while additive quantization reduces it.
- Both quantization schemes reduce the effective data size and introduce an unavoidable additive error term (error floor) due to the 'gap between quantized and full-precision data space'.
- Multiplicative quantization's signal-dependent error decays with the intrinsic signal, preserving spectral structure, whereas additive quantization's constant noise can overwhelm and flatten the spectral tail.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.