Anthropic Research Fellow · AI Security · Agentic RL · Learning Theory
I'm a final-year Ph.D. student at Georgia Tech (ISyE), advised by Prof. Tuo Zhao, and a Research Fellow in the Anthropic Fellows Program. My research aims to make large-scale AI systems reliable and interpretable: quantitative threat models for AI safety and security, RL post-training for LLM agents from environment infrastructure to stable long-horizon training, and theory of diffusion models and deep networks.
How much of a model can an attacker steal under an egress limit? Run 4k+ compress-then-retrain attacks on dense and MoE model ladders and fit scaling laws for the egress limit a lab needs.
Contributed to strands-env, open-source agent infra, and built MCP, user-simulation and terminal-coding environments on a sandbox service, enabling 3k+ concurrent rollouts.
Stable long-horizon agent RL on a 100B+ MoE base, rubric rewards beyond verifiable tasks, and reward learning robust to corrupted preference labels.
Statistical rates for diffusion models on manifolds and for masked discrete diffusion, sample complexity of deep networks set by intrinsic dimension, and a minimal model of edge-of-stability training.
Quantifying how well egress limits protect model weights from exfiltration.
Led agent RL post-training for tool use and terminal coding on a 100B+ MoE base.
Memory-efficient LLM optimization and structured MoE compression.
Full list on Google Scholar