Scientific Director, Aithyra
8 papers at NeurIPS 2025
We present a principled recipe for building graph foundation models that generalize across arbitrary graphs, features, and label spaces.
GradMetaNet is a neural architecture that efficiently processes gradients of other networks by exploiting their symmetries and rank-1 decomposition structure, enabling better learned optimizers, model editing, and loss curvature estimation
We learn non-gradient field dynamics by solving Schrödinger Bridge problem with non-zero reference process drift
we train transferable normalizing flows to sample from peptide Boltzmann distributions up to 8 residues
We propose a diffusion-based approach to sampling from Boltzmann densities based on temperature annealing.
Generate high-performing LoRA parameters from prompts that is unseen in training.
We show the effects of vanishing gradients on GNNs.
We formalize the over-squashing phenomenon in spatiotemporal graph neural networks and analyze how it affects information propagation across the spatial and temporal dimensions.
