Raj S. Shah

My research focuses on the practical use of AI to support human well-being. I approach this through three complementary directions: (1) Technical methods for LLM safety, robustness, and control. I develop techniques such as effective unlearning to protect users, watermarking for authorship verification, and continual learning in real environments. (2) Using pre-trained language models (PLMs) as computational models of human cognition. I use PLMs to better understand human cognition; developing theoretically grounded linking hypotheses, modeling reasoning patterns, and evaluating when and why model behavior aligns with or diverges from human judgments and developmental trajectories. Recently, I have been exploring the developmental alignment of models. (3) Benchmarks and evaluation protocols for practical scenarios and contexts. I design domain-specific evaluations for mental health, social-media visualizations, global representations, clinical documentation, and financial language modeling. These evaluations expose where AI failures have real human consequences. Overall, my research aims to advance value-aligned AI use through technical methods and effective evaluation.