Research Interests

My research lies at the interface of physics, biology, and machine learning. I use ideas from statistical physics to understand how collective behavior and learning emerge in systems composed of many interacting parts—from microbial ecosystems to neural networks and animal behavior. Across these systems, I ask a common question: when can high-dimensional dynamics be explained by a small set of order parameters, statistical laws, or computational strategies? I seek coarse-grained principles that yield quantitative, experimentally testable predictions. Specifically, I develop statistical-physics-inspired modeling, large-scale numerical simulation and deep learning inference tools to invesitage different complex systems and identify the minimal mechanisms that make its behavior understandable.

Representative works

Emergent Organization in Microbial Ecosystems —— Microbial communities can contain thousands of species competing for limited resources. Their composition also changes continually through population turnover, evolution, and horizontal gene transfer. My work investigates how such ecosystems self-ogranize under complexity.

Simplifcity from Complexity: We found that once a community has enough species in it, its overall behavior stops depending on the fine details of which species affects which. All the small differences average out, so a simple model that ignores those details still gets the big picture right. This matters in practice. It means we can predict, and even design, microbiomes without first mapping every single interaction — a task that would be impossible anyway. (Phys. Rev. Lett. 2020; Phys. Rev. E 2021)

Robustness Through Gene Sharing: Real ecosystems never sit still. Species keep evolving, and bacteria pass genes back and forth. We found that a stable community is not one that holds on to the same species, but one that keeps the necessary genes in circulation. To survive, a cell needs the right genes on hand to over natural selections, and those genes are constantly being shared across the community via horizontal gene transfer. Because the genes keep flowing, the community as a whole stays diverse even when individual cells die at a high rate. What defines a community is the pool of genes it holds, not the list of species in it. (PNAS 2025)

Learning with and without Brains

The same puzzle shows up in systems that learn. Learning systems face a problem analogous to that of complex ecosystems: many interacting components must collectively discover useful structure in a vast space of possibilities. I study both artificial and biological learners to understand how computational strategies and behavioral motifs (similar to species in ecology!) emerge, compete, and reorganize during learning.

Generalization versus Memorization in Transformers Language models make a good test case, because we know exactly how they are built and what they were trained on. That lets us watch new abilities appear as they happen. Studying small models as they learn controllable sythestic tasks, we found that whether a model picks up a real rule or simply memorizes answers depends on how varied its training examples are. Surprisingly, a model that is too good at memorizing has a harder time learning to generalize. (arXiv:2604.12151)

Behavioral Strategies for Motor Learning: Animals face a harder problem. They learn by moving through and responding to the real world, where the space of possible behaviors is effectively enormous. I am developing a data-analysis pipeline that reconstructs the full three-dimensional posture of freely moving rats as they acquire new motor skills. Our preliminary analyses suggest that animals navigate this vast space not by inventing entirely new movements, but by selecting from a compact repertoire they already possess. This points to a strong behavioral prior: animals bring pre-existing knowledge about how to move, which constrains the search space and helps them identify successful solutions rapidly. (In preparation.)