I am Yaowen Ye (or Elwin, or 叶耀文), a senior majoring in Computer Science at the University of Hong Kong and an incoming PhD student at UC Berkeley. I am interested in making human oversight of AI systems reliable on hard tasks and efficient for complex outputs. Currently, I am working in Prof. Jacob Steinhardt's group. Before this, I was fortunate to be advised by Prof. Yixin Zhu at PKU Cognitive Reasoning Lab and Prof. Chao Huang at HKU Data Intelligence Lab.
Links:
[X] [Email] [Give me feedback!]
Research Interests
I am interested in making human oversight of AI systems reliable on hard tasks and efficient for complex outputs. Some problems I am thinking about recently are:
- Alignment under unreliable supervision. As humans inevitably make mistakes due to limited capabilities or cognitive and social biases, we need methods that ensure reliable alignment even with systematically flawed supervision.
- Scaling human oversight to interacting AI systems. AI systems will be deployed at massive scale due to their growing capabilities and cost-effectiveness. We need algorithms and tools that help humans efficiently understand the complex outputs of multiple interacting AI systems.
- Scaling laws for human oversight. We need better evaluation of oversight methods, instead of merely measuring accuracy. I aim to develop empirically-grounded scaling laws that help us predict human resource requirements for reliably supervising an AI system on a specific task, hence revealing which oversight methods may remain practical in the future.
In the past, I also worked on cognitive reasoning, intuitive physics, learning on graphs, and recommender systems.
Publications
2025
Auditing Black-Box LLM APIs with a Rank-Based Uniformity Test
Preprint, 2025.
Xiaoyuan Zhu, Yaowen Ye*, Tianyi Qiu*, Hanlin Zhu†, Sijun Tan†, Ajraf Mannan, Jonathan Michala, Raluca Ada Popa, Willie Neiswanger.
[paper]
Iterative Label Refinement Matters More than Preference Optimization under Weak Supervision.
International Conference on Learning Representations (ICLR), 2025. Spotlight presentation.
Yaowen Ye*, Cassidy Laidlaw* and Jacob Steinhardt.
[paper]
2023
Graph Masked Autoencoder for Sequential Recommendation.
ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2023.
Yaowen Ye, Chao Huang and Lianghao Xia.
[paper]
Masked Graph Transformer for Recommendation.
ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2023.
Chaoliu Li, Chao Huang, Lianghao Xia, Xubin Ren, Yaowen Ye and Yong Xu.
[paper]