Ijay Narang

On the machine learning side, I enjoy working on computer vision and reinforcement learning, particularly how models can learn to select and use the visual information a task requires. My interests span applied work, such as post-training models, and theoretical questions about reinforcement learning under various constraints. Recently, I’ve become very interested in understanding how to make reinforcement learning more efficient, particularly by allocating rollout compute more carefully and reusing past experience.

In statistical physics and theoretical computer science, I work on sampling algorithms. A meta question I study is whether sampling is always efficient whenever optimization is. One example is #BIS, the problem of approximately counting independent sets in bipartite graphs: finding a maximum independent set is efficient, while efficient approximate counting and sampling remain open in general. Through the hard-core model, independent sets become configurations of a statistical physics system. I’m interested in how phase transitions and the geometry of these configurations help explain computational barriers, and how that understanding can guide the design of sampling algorithms.