About
I am a Senior Research Scientist at Google Cloud AI Research, where I
work on LLM agents and agentic RL, with a focus on long-horizon coding and machine-learning-engineering (MLE)
agents.
My recent work includes MARS, an MLE agent that
reached #1 on the MLE-Bench leaderboard, and test-time scaling methods such as
SETS and TUMIX.
I also contribute to frontier model post-training, building SFT data generation pipelines and designing RL rewards for
ML-engineering tasks. I obtained my Ph.D. degree from the Computer Science
Department at the University of Wisconsin-Madison,
where I was co-advised by Prof. Yingyu Liang and
Prof. Somesh Jha
and supported by the Center for Trustworthy Machine Learning
(CTML).
My thesis research was on trustworthy machine learning, spanning adversarial robustness, robust interpretability, and
out-of-distribution detection.
I obtained my Bachelor's degree in Computer Science from Shanghai Jiao Tong University (SJTU).
News
05/25/2026: We released ScientistOne, an agent towards human-level autonomous research via chain-of-evidence.
03/19/2026: CoDA, our agentic system for collaborative data visualization, is now open-sourced under the
google-research organization.
02/17/2026: Our MLE agent MARS achieved the top-1 position on the MLE-Bench
leaderboard.
11/06/2025: New publication on Google Research Blog: DS-STAR: A state-of-the-art versatile data science agent.
08/01/2025: New publication on Google Research Blog: MLE-STAR: A state-of-the-art machine learning engineering agent.
01/18/2024: New publication on Google Research Blog: Introducing ASPIRE for selective prediction in LLMs.
01/20/2023: Our paper The Trade-off between Universality and Label Efficiency of Representations from Contrastive Learning was accepted by ICLR 2023, as a Spotlight pressentation.
01/20/2023: Our paper Is Forgetting Less a Good Inductive Bias for Forward Transfer? was accepted by ICLR 2023. This work was done while I was interning at DeepMind.
02/28/2022: Our paper Revisiting Adversarial Robustness of Classifiers With a Reject Option received Best Paper Award at
AAAI Workshop 2022.
11/25/2021: Wrote a blogpost about our paper
Detecting Errors and Estimating Accuracy on Unlabeled Data with Self-training Ensembles,
which was accepted by NeurIPS 2021.
06/18/2021: Our paper ATOM:
Robustifying Out-of-distribution Detection Using Outlier Mining was accepted by ECML 2021 (Acceptance Ratio: 21%).
10/31/2019: Wrote a blogpost about our paper
Robust Attribution Regularization,
which was accepted by NeurIPS 2019.