I am currently a Postdoctoral Researcher in Prof. Dawn Song’s group at UC Berkeley. Before that, I earned my Ph.D. in Computer Sciences from the University of Wisconsin–Madison, advised by Prof. Sharon (Yixuan) Li.
My current research focuses on modern LLM agents’ training and evaluation philosophy:
Can AI agents complete the kind of long-horizon work that creates real economic value?
A living benchmark for professional workflows with verifiable outcomes. It spans 55 sub-industries and more than 1,500 tasks contributed by over 300 domain experts, evaluating whether an agent can deliver the work rather than merely answer a question about it.
Highlights: OpenAI’s GPT‑5.6 release; 50+ media coverage; 100K+ downloads in one month.
How can reinforcement learning unlock algorithms that a base model cannot solve at all?
DELTA isolates fully out-of-distribution coding families and reveals a striking phase transition: after a long near-zero-reward plateau, a staged reward recipe can push models from pass@K=0 to reliable solutions.
Highlights: ICLR 2026; featured by Berkeley RDI; covered by AI Era (新智元).
Can LLMs reason outside the box in mathematics?
A controlled benchmark that separates three kinds of generalization—exploratory, compositional, and transformative. Frontier models improve on harder familiar problems, but genuine strategy shifts remain elusive.
Highlights: NeurIPS 2025; featured by Ai2; covered by AI Era (新智元).
A unified view of hallucination through subsequence associations.
The framework explains hallucination as a competition between faithful and misleading associations encoded by a transformer. A tracing algorithm identifies causal subsequences across randomized contexts and connects them to evidence in the training corpus.
Highlights: NeurIPS 2025; theoretical framework and causal tracing; open-source code.