Dec 2025	[🚨 New paper] We depeloped MSPT - parallelized multi-scale attention method based on hierarchical partitioning of data. It is incredibly fast and achieves SOTA performance on multiple PDE tasks.
Dec 2025	[🥳 Paper accepted] Our submission Adaptive Mesh Quantization for Neural PDE Solvers was accepted to TMLR! We suggest a data-driven way of quantizing message-passing neural networks inspired by speculative decoding.
Oct 2025	[🚨 Workshop paper] We introduce Conditional Clifford-Steerable convolution, where we condition the convolutional operator with auxiliary information, achieving highly expressive, equivariant, and efficient framework which we apply to PDE modelling. Accepted to AI4Science Workshop at NeurIPS 2025.