Ankur Dahiya

Hi, I'm Ankur. I'm an undergrad studying Data Science & Artificial Intelligence at IIT Roorkee, and a core member of the Vision and Language Group (VLG).

My research sits at the intersection of understanding and improving large-scale models. I work on mechanistic interpretability — figuring out the algorithms neural networks actually learn inside — and on RL alignment, building training methods that go beyond simple scalar rewards to produce models that do what we actually want.

I also spend a lot of time on model pruning (we got a paper into AAAI on evolutionary pruning of LLMs) and on the compute infrastructure side — the pipeline optimization, cluster health-checks, and training efficiency work that nobody talks about but everybody needs.

Outside of research, I like long walks, overthinking font choices, and occasionally submitting to the void that is academic peer review.