Weijie Su, a statistics professor at the University of Pennsylvania's Wharton School, announced on X that he has officially joined OpenAI during his sabbatical from Wharton. Su, who earned his PhD in statistics from Stanford and his bachelor's from Peking University, said he's returning to the Bay Area after a decade to work on AI model training. He also revealed that he has been promoted to full professor in Wharton's Department of Statistics and Data Science.
https://twitter.com/weijie444/status/2060604060362014803
Su's work spans statistics, machine learning optimization, and the theoretical safety of large language models. In 2026, he won the COPSS Presidents' Award, one of the highest honors in international statistics. The award citation praised his foundational contributions to the statistical foundations of generative AI — including watermarking mechanisms for LLMs, human preference alignment, and ranking theory — as well as his work on differential privacy methods for the 2020 US Census and his contributions to deep learning theory, convex optimization, and high-dimensional statistical inference.
On the research front, Su has made several strides in reshaping the statistical underpinnings of large models. For watermark detection, he formulated watermarking as a statistical hypothesis testing problem and proposed the truncated goodness-of-fit test (Tr-GoF), which improves robustness against modified or cropped text without relying on prior knowledge of human editing processes or exact probabilities. On preference alignment (RLHF), Su pointed out that traditional RLHF's ranking-based optimization can easily lead to "preference collapse" — ignoring minority preferences and causing reward distributions across different prompts to become uniform. To preserve response diversity, his team derived a "preference matching" regularization scheme by solving ordinary differential equations (ODEs), maximizing human rewards while avoiding homogenized responses. They also explored a game-theoretic approach called "Nash learning from human feedback" (NLHF) as an alternative to traditional scalar reward alignment.
Su earned his bachelor's from Peking University's School of Mathematical Sciences, part of what media and scholars call the school's "Golden Generation II" (also referred to as "Platinum Generation"). Compared with the pure-math-focused "Golden Generation I" (class of 2000, including mathematicians like Yun Zhiwei, Zhang Wei, and Xu Chenyang), the class of 2007 represented by Su exhibits a broader interdisciplinary profile, extending mathematics into statistics, data science, and artificial intelligence.