Assistant Professor, Technion - Israel Institute of Technology, Technion
2 papers at NeurIPS 2025
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 propose to treat LLM activations as images and detect hallucinations more accurately (and efficiently) across LLMs with a vision-inspired architecture.
