* indicates equal contribution or alphabetic order. The arXiv version is preferred for the most up-to-date content.
Yuxin Ma, Adir Dayan, Yam Eitan, Haggai Maron, Soledad Villar
arXiv preprint.We study size generalization in weight-space learning. We introduce graph metanetworks that predict properties of neural-network weights and generalize out of distribution to input networks with much larger widths.
Yuxin Ma, Nan Chen, Mateo Díaz, Soufiane Hayou, Dmitriy Kunisky, Soledad Villar
ICML 2026.We propose a theoretically grounded model-upscaling method that uses a trained smaller neural network to initialize a larger one. The algorithm is based on dynamic equivalences between models of different widths and the μP infinite-width scaling limit, enabling zero-shot hyperparameter transfer.
Eitan Levin*, Yuxin Ma*, Mateo Díaz, Soledad Villar
NeurIPS 2025 (Spotlight).We introduce a mathematical framework for understanding when machine learning models generalize across input dimensions, such as graph neural networks across graph sizes, Deep Sets across set sizes, and transformers across sequence lengths.
Yuxin Ma, Dmitriy Kunisky
NeurIPS 2025 (Spotlight).We study a new class of spectral algorithms for low-rank estimation based on tunable nonlinear deformations of the observed matrix. The deformation can be chosen by black-box optimization or learned from data using neural networks.