LLM agents and agentic RLs, including environment and task constructions, specialized model training, harness and evaluation designs
Agent Safety and Security, including lost-of-control alignment, monitoring, red-teaming and guardrails
At Meta, I contribute to post-training and evaluation of the Muse models on cybersecurity and coding, as well as system safety for the Muse-related products.
My previous works studied the trustworthiness of DRL and DNN via statistical modeling (e.g., Deep GP, Gaussian process
models) and learning (e.g., variational inference, MCMC, empirical bayes),
and designing DNNs for security applications with a focus on noisy learning and OOD.
Honors and Awards
Academic Senate Faculty Research Award, UCSB, 2025
