Pragya Srivastava

I am a researcher at Google DeepMind, working on Robust and Sample Efficient Reinforcement Learning for post-training Gemini with Prof. Doina Precup. Dr. Karthikeyan Shanmugam, and Dr. Aravindan Raghuveer.

Previously, I was a Research Intern at VITA, UT Austin, where I worked on the theoretical analysis of Structured State Space Models. I also had the pleasure to work with Dr. Amit Sharma and Dr. Amit Deshpande at Microsoft Research, where I worked on In-context Learning and OOD detection in Reward models. Before this, I completed my undergrad from Indian Institute of Technology, Delhi, with a major in Engineering Physics. During this time, I was a research intern at UCSD, advised by Prof. Pengtao Xie, and at the Ermon Group, Stanford University.

Email  /  CV  /  Google Scholar  /  X  /  Github

Research

My research focuses on identifying and characterizing the fundamental failure modes of existing AI alignment methods. I am keen in developing principled approaches to enable reasoning models to explore effectively through interventional feedback, and generalize robustly out-of-domains. Broadly, I aim to bridge the gap between theoretical learning algorithms and practical deployment to build models that are highly capable yet fundamentally safe.

We study the effects of covariance sphering and welfare-based aggregation to convert reward vectors into scalars for policy alignment. This resolves signal redundancy and reward hacking.

Developing robust reward models with reduced reliance on spurious attributes and higher sensitivity to causal rubrics.

We systematically study monitoring for out-of-distribution alignment failures of LLMs using guard models and OOD detectors.

Dynamic alignment method using energy-based OOD scoring to identify misjudgments and refine both policy and reward model.

We reveal two core limitations of state space models—recency bias and over-smoothing—and propose a polarization technique improving long-range memory and stability.

We introduce a task-specific metric, NICE (Normalized Invariability to Choice of Examples), that quantifies task learnability and guides whether to optimize instructions or ICEs.

LLMs perform well on simple tasks but degrade sharply with table complexity, requiring deeper reasoning and multi-row inference.