My approach to research is deeply rooted in the principles of experiment-driven research, careful engineering, and efficiency. I believe that the best way to advance our understanding of complex systems, like large language models and reinforcement learning, is through rigorous experimentation. In my work, I strive to design and conduct out-of-the-box experiments that test our current understanding and unlock avenues to improved performance, while maintaining statistical and empirical rigor.
This is vital, because reinforcement learning has long struggled with the combination of sloppy, underpowered empirical practices and an ideological attachment to theoretically-tractable methods that do not perform competitively outside of small regimes. In my experience, “theoretically unprincipled” (read: insane) ideas, such as aggressively resetting agent parameters, can in reality allow revolutionary improvements in performance, so long as our experiments allow us to find and evaluate them reliably.
I’m also a firm believer in the power of careful engineering. When performing empirical research, every additional training run counts towards achieving a statistically meaningful result. But efficiency, to me, is not just about speed. It’s about making the most of our resources, whether that’s computational power, training data, or our own time. With the right engineering, thorough empirical research is possible even with limited academic resources.