Hi, I am Diksha :)

I study how neural networks solve complex cognitive tasks, and diagnose when and how they fail. My interests lie at the intersection of LLM training dynamics, alignment, and interpretability.

I am currently a senior research fellow at Pivotal, working with Stefan Heimersheim on the continuity of LLM activation manifolds and what that implies for manifold-aware steering and probing. Previously, I worked on red-teaming unlearning methods for open-weight safety.

Before that, I was a senior research fellow at the Sainsbury Wellcome Centre at University College London, with the Behrens and Mrsic-Flogel labs, studying compositional representations and their learning algorithms in artificial and biological neural networks. During my PhD with Carlos Brody at Princeton University, I developed ablation-aware methods to train RNNs on cognitive tasks to improve capability attribution, and better models to track systematic biases in decision-making and predict previously unpredictable failures, working from neural recordings and causal experiments in rodents playing carefully crafted structured games.