Lun Wang
Make agents reliable at OpenAI. Previously staff research scientist at Google DeepMind on Gemini post-training. I received my PhD in computer science from UC Berkeley, advised by Prof. Dawn Song in 2022. Previously, I earned my Bachelor's degree with honors in computer science from Peking University in 2018.
Blog
-
AI 安全和三体问题 · AI Safety and the Three-Body Problem
When alignment and control both fall short, the Deterrence Era suggests a third path: don't predict what it wants, change the consequences it faces.
-
Your Evals Will Break and You Won't See It Coming
Why LLM evaluations may fail exactly when models undergo qualitative phase transitions.
Industry Projects
Selected Publications
-
OpenSage: Self-programming Agent Generation Engine.
ICML 2026: Forty-Third International Conference on Machine Learning. -
DevOps-Gym: Benchmarking AI Agents in Software DevOps Cycle.
ICLR 2026: The Fourteenth International Conference on Learning Representations. -
Revisit Micro-batch Clipping: Adaptive Data Pruning via Gradient Manipulation
ICLR 2025: The Thirteenth International Conference on Learning Representations. -
Why Is Public Pretraining Necessary for Private Model Training?
ICML 2023: Fortieth International Conference on Machine Learning. -
Secure Federated Correlation Test and Entropy Estimation.
ICML 2023: Fortieth International Conference on Machine Learning. -
Byzantine-Robust Federated Learning with Optimal Statistical Rates and Privacy Guarantees.
AISTATS 2023: Artificial Intelligence and Statistics. -
PRIVGUARD: Privacy Regulation Compliance Made Easier.
Usenix Security'22: the 31th Usenix Security Symposium. -
Differentially Private Fractional Frequency Moments Estimation with Polylogarithmic Space.
ICLR 2022: The Tenth International Conference on Learning Representations. -
DUET: An expressive higher-order language and linear type system for statically enforcing differential privacy.
OOPSLA'19: Proceedings of the ACM on Programming Languages. Distinguished Paper Award. -
Towards practical differentially private convex optimization.
Oakland'19: 2019 IEEE Symposium on Security and Privacy.