Hershraj Niranjani

About

Hi! I'm Hershraj. I'm broadly interested in how we can get large models to reason effectively in structured domains. At BAIR, I worked on LLM-guided program synthesis -- figuring out how to get language models to search well over massive spaces of possible programs. I've also published work on multi-agent coordination, looking at how LLM systems can autonomously discover their own behavioral norms rather than relying on hand-designed ones. I've also spent time exploring physics-informed neural networks, timeseries forecasting, and interpretability. I tend to gravitate toward problems where the structure of the domain matters as much as the model itself.

Selected Work