Daogao Liu

I am a Research Scientist at Google Research, where I work on post-training for Gemini. I lead evaluation for Gemini Thinking and drive RL and SFT data curation, synthesis, and evaluation methodology across knowledge work, web development, reasoning, cybersecurity, and software engineering. I also design and scale multi-agent and recursive self-improvement (RSI) pipelines to push frontier agentic capabilities. More recently, I have started working on pretraining, drawing on my background in continuous optimization.

Before that, I was a visiting postdoctoral researcher at Google (2024–2025), where I pivoted from pure theory to large-scale empirical LLM research. During that time I led Urania (COLM 2025; Google Research blog), a differentially private framework for extracting insights from chatbot conversations that has since been productized and is used internally at Google; was a core contributor to VaultGemma, a 1B-parameter open LLM pretrained with formal differential privacy guarantees; and contributed reasoning training data to the Gemini Deep Think model that achieved gold-medal standard at the IMO.

I received my PhD in Computer Science from the University of Washington in 2024, advised by Prof. Yin Tat Lee, and a B.S. in Mathematics and Physics from Tsinghua University in 2020. My PhD research was on differentially private optimization, stochastic optimization, and algorithm design, with 30+ papers at venues including NeurIPS, ICML, ICLR, COLT, STOC, FOCS, and SODA. I was an Apple Scholar in AI/ML in 2024.