Research

I am interested in how learning objectives and training choices shape the behavior of intelligent systems and in how those changes can be studied clearly and efficiently. This page describes questions I am actively exploring; completed work is listed in Experience and Publications.

Understanding LLM Behavior

My main research interest is understanding how training and alignment methods change a language model’s behavior. I am particularly interested in uncertainty, calibration, and the trade-offs introduced by supervised fine-tuning, Direct Preference Optimization (DPO), and reinforcement learning.

The questions I keep returning to are concrete ones: what changes when a model is trained to imitate examples rather than satisfy preferences? How do different objectives shift the probability assigned to answers, reasoning traces, and failure modes? And which measurements can reveal those changes beyond an aggregate task score?

This direction grows out of my work on LLM uncertainty quantification and uncertainty-informed DPO at MBZUAI.

Learning Through Controlled Environments

I am also studying reinforcement learning from first principles. Games are a useful setting for this because the environment is structured, policies can be examined repeatedly, and it is easier to trace a behavior back to the training choices that produced it.

I am exploring a Pokémon TCG bot as a personal learning environment for these ideas. It is an ongoing study rather than a completed project.

Making Experiments Practical

Finally, I care about the engineering work that makes research easier to run, inspect, and reproduce: efficient training loops, memory-conscious inference, and implementations that make experiments affordable enough to iterate on.

I see this as part of the same research process. Clear experiments make it easier to understand what a learning method is actually doing, rather than only observing its final score.