News

Oct 06, 2026 New preprint with Clarissa Lauditi and Cengiz Pehlevan — A Solvable Model of Adaptive Learning Rate Rescaling, where we use DMFT to study how normalized SGD accelerates training, where that acceleration breaks down at the edge of stability, and what it means for trading off width vs. batch size.
Sep 27, 2026 I will present Phase Transitions in Attention in NeurIPS, see you in Sydney!
Jun 16, 2026 New preprint out — Phase Transitions in Attention, where we work out a Bayesian theory for how copy heads emerge, and find a an interesting phase diagram.
Sep 01, 2025 Started my PhD in applied mathematics at Harvard, joining the Pehlevan Group.