Hunar Batra

Hunar Batra

PhD Computer Science, University of Oxford

I work on AI interpretability and control: understanding, monitoring and controlling LLM behaviours, actions and latent dynamics.

About

I'm a PhD Computer Science student at the University of Oxford, working on AI interpretability and control at the PIXL Lab, supervised by Prof. Ronald Clark.

My research explores understanding, monitoring and controlling large language models' behaviours, actions and latent dynamics, from understanding emergence and detecting signs of misalignment and reward hacking, to surfacing models' internal computations in reasoning faithfully, white-box auditing of models, safety training and scalable oversight.

In the past, I was a Research Engineer (contract) at Anthropic working on evals, human data and post-training LLMs for generating robust reasoning, at METR contributing to AI R&D evals, at the UK AI Security Institute building multi-agent safety evaluations, and at Cadenza Labs red-teaming models and training probes for detecting deception. Before these, I earned an MSc in Advanced Computer Science at Oxford, where my research focused on training protein language models to predict SARS-CoV-2 mutations.

Publications

Google Scholar 1,611 citations

Blog

Chain of Thoughts

Notes and thoughts.

Coming soon