Florian Wolf

Hi, I am a PhD student in Applied and Computational Mathematics (ACM) at Caltech, working with Prof. Andrew Stuart and supported by the Kortschak Scholars Program Fellowship.

My research interests are centered around

• Optimization and Reinforcement Learning
• (Multimodal) Representation Learning

Email  /  Scholar  /  Github  /  LinkedIn

News:

• 10/2026, Oral@EWRL I will present our paper Global Optimality for Constrained Exploration via Penalty Regularization in Lille, France
• 08/2026, new preprint on arxiv: Expressivity in Multimodal Contrastive Learning

Thesis / Project supervision:

If you're passionate about research and interested in working with me, I'm always happy to hear from motivated students. Please reach out with a summary of your research interests, along with your up-to-date CV and academic transcript.

We develop a universal approximation theory for multimodal contrastive learning. While classical CLIP is universal for two modalities, pairwise extensions cannot capture general interactions among three or more modalities. We introduce Hadamard-CLIP, which restores joint universality while retaining precomputable embeddings and efficient retrieval.

While maximum-entropy exploration is central to RL, real-world applications are often hindered by safety constraints and a lack of additive structure. We propose Policy Gradient Penalty (PGP), a single-loop method that enforces general occupancy-measure constraints via quadratic penalty regularization. By leveraging hidden convexity, we provide the first global last-iterate convergence guarantees for this non-convex setting.

First algorithms with provable global guarantees for constrained non-convex optimization via hidden convexity. Our methods bypass constraint qualifications, handle hidden convex equality constraints, and work directly with gradient oracles in the non-convex space. Applications include safe control and reinforcement learning, where we establish global optimality and oracle complexities matching unconstrained hidden convex optimization.

AE+SINDy-C: a data-efficient, interpretable, and scalable Dyna-style Model-Based RL framework for PDE control, combining SINDy-C with autoencoders for dimensionality reduction. Applied to the 1D Burgers and 2D Navier-Stokes equations, the method enables fast rollouts, reduces environment interactions by up to 10x, and yields an interpretable latent dynamics model, outperforming a model-free baseline.

Bayesian spatio-temporal product partition model to cluster PM2.5 air quality data from multiple monitoring stations in Northern Italy, capturing both spatial and temporal patterns. The model outperforms a spatial-only baseline in predictive performance, offering smoother and more insightful pollution trend analysis.

Automated curriculum generation with massively parallel RL learns a spherical pendulum tracking controller, leveraging the task's non-Euclidean structure for faster convergence, higher performance, and successful sim-to-real transfer.

The website is based on the code from Jon Barron