# New Preprint Attacks the Arithmetic of Low-Precision Formats, Not Just Their Storage

CurveFP proposes a codebook family whose products stay inside the format, targeting the error that microscaling four-bit types introduce during multiplication rather than during rounding.

- Published: 2026-08-12T06:11:01.127Z
- Canonical: https://polylog.news/ai/2026-08-12/new-preprint-attacks-the-arithmetic-of-low-precision-formats
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [arXiv](https://arxiv.org/abs/2608.10010), [Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4), [Google Research](https://research.google/research-areas/)

A preprint posted to arXiv, CurveFP, makes an argument that most quantization work has sidestepped. Low-precision datatypes are usually designed to minimize the error in representing a single number, while the arithmetic error that arises w…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-08-12/new-preprint-attacks-the-arithmetic-of-low-precision-formats (subscription information: https://polylog.news/pricing).