Dissecting Quantization Error: A Concentration-Alignment Perspective
Marco Federici, Boris van Breugel, Paul Whatmough, Markus Nagel · 2026
Read Dissecting Quantization Error: A Concentration-Alignment Perspective by Marco Federici, Boris van Breugel, Paul Whatmough, Markus Nagel (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-04359.txt.
Description
Federici and colleagues explain why rotation style transforms help quantized models. Quantization boosts efficiency but usually costs accuracy, for unclear reasons. The quantization noise ratio splits into concentration plus alignment of main directions. Prior transforms target concentration alone, but better alignment can cut error further. Block CAT transforms use a small calibration set to lift both factors at once. At 4-bit precision CAT matches or beats prior transform methods on several large language models.
Catalog record
| CREATOR | Marco Federici, Boris van Breugel, Paul Whatmough, Markus Nagel |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46706 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.04359