LATMiX: Learnable Affine Transformations for Microscaling Quantization of LLMs

Ofir Gordon, Lior Dikstein, Arnon Netzer, Idan Achituve, Hai Victor Habi
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:36315-36346, 2026.

Abstract

Post-training quantization (PTQ) is a widely used approach for reducing the memory and compute costs of large language models (LLMs). Recent studies have shown that applying invertible transformations to activations can significantly improve quantization robustness by reducing activation outliers; however, existing approaches are largely restricted to rotation or Hadamard-based transformations. Moreover, most studies focused primarily on traditional quantization schemes, whereas modern hardware increasingly supports the microscaling (MX) data format. Attempts to combine both showed severe performance degradation, leading prior work to introduce assumptions on the transformations. In this work, we take a complementary perspective. First, we provide a theoretical analysis of transformations under MX quantization by deriving a bound on the quantization error. Our analysis emphasizes the importance of accounting for both the activation distribution and the underlying quantization structure. Building on this analysis, we propose LATMiX, a method that generalizes outlier reduction to learnable invertible affine transformations optimized using standard deep learning tools. Experiments show consistent improvements in average accuracy for MX low-bit quantization over strong baselines on a wide range of zero-shot benchmarks, across multiple model sizes.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-gordon26b, title = {{LATM}i{X}: Learnable Affine Transformations for Microscaling Quantization of {LLM}s}, author = {Gordon, Ofir and Dikstein, Lior and Netzer, Arnon and Achituve, Idan and Habi, Hai Victor}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {36315--36346}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/gordon26b/gordon26b.pdf}, url = {https://proceedings.mlr.press/v306/gordon26b.html}, abstract = {Post-training quantization (PTQ) is a widely used approach for reducing the memory and compute costs of large language models (LLMs). Recent studies have shown that applying invertible transformations to activations can significantly improve quantization robustness by reducing activation outliers; however, existing approaches are largely restricted to rotation or Hadamard-based transformations. Moreover, most studies focused primarily on traditional quantization schemes, whereas modern hardware increasingly supports the microscaling (MX) data format. Attempts to combine both showed severe performance degradation, leading prior work to introduce assumptions on the transformations. In this work, we take a complementary perspective. First, we provide a theoretical analysis of transformations under MX quantization by deriving a bound on the quantization error. Our analysis emphasizes the importance of accounting for both the activation distribution and the underlying quantization structure. Building on this analysis, we propose LATMiX, a method that generalizes outlier reduction to learnable invertible affine transformations optimized using standard deep learning tools. Experiments show consistent improvements in average accuracy for MX low-bit quantization over strong baselines on a wide range of zero-shot benchmarks, across multiple model sizes.} }
Endnote
%0 Conference Paper %T LATMiX: Learnable Affine Transformations for Microscaling Quantization of LLMs %A Ofir Gordon %A Lior Dikstein %A Arnon Netzer %A Idan Achituve %A Hai Victor Habi %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-gordon26b %I PMLR %P 36315--36346 %U https://proceedings.mlr.press/v306/gordon26b.html %V 306 %X Post-training quantization (PTQ) is a widely used approach for reducing the memory and compute costs of large language models (LLMs). Recent studies have shown that applying invertible transformations to activations can significantly improve quantization robustness by reducing activation outliers; however, existing approaches are largely restricted to rotation or Hadamard-based transformations. Moreover, most studies focused primarily on traditional quantization schemes, whereas modern hardware increasingly supports the microscaling (MX) data format. Attempts to combine both showed severe performance degradation, leading prior work to introduce assumptions on the transformations. In this work, we take a complementary perspective. First, we provide a theoretical analysis of transformations under MX quantization by deriving a bound on the quantization error. Our analysis emphasizes the importance of accounting for both the activation distribution and the underlying quantization structure. Building on this analysis, we propose LATMiX, a method that generalizes outlier reduction to learnable invertible affine transformations optimized using standard deep learning tools. Experiments show consistent improvements in average accuracy for MX low-bit quantization over strong baselines on a wide range of zero-shot benchmarks, across multiple model sizes.
APA
Gordon, O., Dikstein, L., Netzer, A., Achituve, I. & Habi, H.V.. (2026). LATMiX: Learnable Affine Transformations for Microscaling Quantization of LLMs. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:36315-36346 Available from https://proceedings.mlr.press/v306/gordon26b.html.

Related Material