Publications

μpscaling small models: Principled warm starts and hyperparameter transfer

Yuxin Ma, Nan Chen, Mateo Díaz, Soufiane Hayou, Dmitriy Kunisky, Soledad Villar

Preprint (2026).

We propose a μP-based model upscaling method that allows hyperparameter transfer.

model upscaling hyperparameter transfer mup tensor program training dynamics infinite-width limit

[arXiv(preferred)] [Code]

On transferring transferability: Towards a theory for size generalization

Eitan Levin^*, Yuxin Ma^*, Mateo Díaz, Soledad Villar

NeurIPS 2025 (Spotlight).

We study the properties that make machine learning models generalize their performance across input dimensions.

transferability size generalization graph neural networks equivariant machine learning any-dimensional learning

[arXiv(preferred)] [NeurIPS] [Code] [Poster]

Nonlinear Laplacians: Tunable principal component analysis under directional prior information

Yuxin Ma, Dmitriy Kunisky

NeurIPS 2025 (Spotlight).

We study a new class of spectral algorithms for low-rank estimation that involve a tunable nonlinear deformation of an observed matrix. The tunable component can be selected using methods such as black-box optimization or learned from data using neural networks.

principal component analysis random matrix theory spiked matrix models low-rank estimation

[arXiv(preferred)] [NeurIPS] [Code] [Poster]