Hi, I'm Deven Choudhary. I work on mechanistic interpretability.
Mostly I take language models apart to see which computations they reuse and which they quietly rebuild from scratch — whether a sentiment circuit that works in English survives the trip into Hindi (it mostly doesn't), or whether a model's existing "disgust" direction gets drafted in to do other, uglier jobs. Less benchmark-chasing, more figuring out what's actually inside.
Selected work