Research Director, Netflix
5 papers at NeurIPS 2025
Adaptive instrumental variable design plus a multiply-robust estimator yields efficient, unbiased ATE estimates in sequential settings.
We introduce a theoretically-grounded distributional RL algorithm for LLM post-training that demonstrates improvement upon prior work on both synthetic and mathematical reasoning tasks.
We introduce GST-UNet, a neural framework for valid causal inference from spatiotemporal observational data with time-varying confounding and complex spatiotemporal dependencies.
