Hi! I'm Zhiwei, a fourth-year Ph.D. candidate at the University of Michigan, Ann Arbor, gratefully advised by Prof. Wei Hu and Prof. Yixin Wang. I'm interested in the science of training foundation models and agents, from scaling laws and model architecture to data and training recipes for agentic models. This summer, I'm interning at Google Cloud, working on credit assignment for CLI agents, hosted by Haichao Yu. Previously, I interned at Google Research, working on model architecture, hosted by Hrayr Harutyunyan.
Education
Publications & Preprints
Terminal-Bench: A Benchmark for Agents on Hard, Realistic Tasks in Command Line Interfaces
ICLR 2026.
Transformers with Improved Horizontal Information Flow
Submitted as Google Patents.
Benign Overfitting in Single-Head Attention
NeurIPS 2025.
Inference of Dependency Knowledge Graph for Electronic Health Records
Journal of the Royal Statistical Society, Series B, 2025
ARCH: Large-scale Knowledge Graph via Aggregated Narrative Codified Health Records Analysis
Journal of Biomedical Informatics, 2025
Teaching
STATS 500: Statistical Learning
Fall 2024Graduate Student Instructor, University of Michigan
Graded homework and conducted office hours
STATS 413: Applied Regression Analysis
Winter 2024Graduate Student Instructor, University of Michigan
Taught weekly lab classes, graded homework, and conducted office hours
STATS/DATASCI 451: Bayesian Data Analysis
Fall 2023Graduate Student Instructor, University of Michigan
Graded homework and conducted office hours
STATS 250: Introduction to Statistics and Data Analysis
Winter 2023Graduate Student Instructor, University of Michigan
Taught weekly lab classes, graded homework and lab, and conducted office hours
Real Analysis
Spring 2021Teaching Assistant, USTC
Taught weekly lab classes and graded homework
Service
- Conference Reviewer: ICML (2024, 2025), ICLR (2025, 2026), L4DC (2025)
- Workshop Reviewer: NeurIPS M3L (2024)