Hello! I am a final-year PhD candidate at the Paul G. Allen School of Computer Science & Engineering, University of Washington. I am fortunate to be co-advised by Hannaneh Hajishirzi and Noah Smith. I am also a student researcher at Ai2. I have previously interned at Meta AI and Microsoft Research Asia. Prior to UW, I did a master at Peking University and an undergraduate at Shanghai Jiao Tong University.

I work broadly on natural language processing, machine learning, and artificial intelligence. In particular, I enjoy studying fundamental data challenges in AI development and how to advance AI through data creation. My work has led to the unification of NLP tasks with instruction tuning (e.g., Super-NaturalInstructions), pioneered the use of synthetic data for language model training (e.g., Self-Instruct), and scientified the development of fully open language models (e.g., OLMo and Tülu). I believe that data serves as the foundation that defines the behavior and upper limits of AI models. It also provides an effective, interpretable, and beneficial ground to advance AI as a community. I have been thinking about these topics recently:

• Data for generality. The major focus of AI is progressing from single-purpose applications to a new stage aiming to build better general-purpose models. Optimizing towards this goal requires data beyond traditional paradigms and with a unified view of tasks, modalities, and domains. This poses many opportunities for data creation and calls for data-centric algorithms in achieving the right balance among diversity, speciality, quantity, quality, efficiency, and many other factors.
• Synthetic data. One promising direction for creating data is to automatically synthesize it using models, rules, or environments. Synthetic data not only enables easy adaptation of models to specific scenarios, but also opens up potentials for model self-improvement. Many active topics are also very related, such as data augmentation, AI feedback, and the exploration of actions in RL. Research in this area is still in full swing and has many open questions to be answered.
• Open science of language models. Openness is critical for collaboration, innovation, and accountability, while today's frontier models are trending oppositely. I am passionate about building fully open alternatives, to inform the public about the science behind these models and their potential risks. This also drives studies in model training methods, their interplay with data, and how we can improve them with the open and research community.