I work at the interface of theoretical physics and machine learning. Most strongly correlated systems are intractable; I look for the exceptions — the ones where correlations are driven by a handful of hidden variables.
Where statistical mechanics meets generative modeling and AI safety. Why a VAE is structurally a finite-size mean-field model, and what that costs when the data isn't — and why jailbreaking an aligned model looks like a particle escaping a trap.
Non-interacting particles coupled only by a shared reset event — enough to correlate a gas that would otherwise never meet. Exact results for extremes, gaps, and order statistics follow from conditioning on a single hidden variable.
Dyson Brownian motion of eigenvalues, reset simultaneously. Stationary density, extreme-eigenvalue statistics, and the crossover between repulsion- and reset-dominated regimes.
A jailbreak is a particle escaping a trap.
Best-of-N jailbreaking breaks an aligned model by brute force: draw N random augmentations of an unsafe prompt, sample M completions of each, and keep whatever gets through. Its success rate has been reported to follow a power law in N. We think it doesn't. The drifting exponent is a finite-size artifact of the adversarial dataset and relatively narrow fitting windows.
We introduce a physically motivated barrier model with stochastic dynamics. Each prompt has a baseline safety level, and each augmentation adds a random, thermally activated barrier. Four interpretable numbers then fix the entire (N, M) attack surface. They extrapolate from N ≤ 100 to N = 104, collapse five models onto one scaling function, and predict attacks at temperatures they were never fitted on.
Statistical physics turned on AI safety: best-of-N jailbreaking as thermally activated escape over random barriers. Four interpretable numbers predict the full attack surface, far beyond the sampling budgets they were fitted on.
With Max Welling and Vincenzo Vitelli — a capacity bound showing that any successful VAE decoder is a finite-size mean-field factorization, so the microscopic theory can be read off the trained network. Applied to salamander retinal recordings.
Presented Variational auto-encoders are finite-size mean-field approximators at the Schmidt AI in Science Speaker Series, University of Chicago.
A Brownian gas whose collective reset is triggered by the first particle to reach a threshold. Now published in EPL 153, 31002 (2026).
With Gabriele de Mauro, Satya N. Majumdar and Grégory Schehr — how non-Poissonian reset protocols reshape the correlation structure of the resetting gas. Now published in Phys. Rev. E 113, 014120 (2026).
Started as a Research Scholar across the Physics and Computer Science departments, working with Vincenzo Vitelli — supported by the Eric & Wendy Schmidt AI in Science fellowship.
Can we learn from our mistakes? A note on my belief in our moral and ethical responsibility as scientists with regard to our discoveries.
293 citations · h-index 10 — Google Scholar, 2026-09-28
I'm happy to hear from anyone working on correlated noise, exact results, or the statistical physics of learning. Email is fastest.