About Me

I am a Ph.D. Candidate at the School of Information of University of Michigan advised by Professor Paramveer Dhillon. Also, I am honored to work closely with Professor Yixin Wang. Before that, I earned my B.S. degree in Computer Science from The Ohio State University, transferred from the School of Informatics at Xiamen University. I've also been lucky to work with Professor David Jurgens at Blablablab.

I am broadly interested in causal inference, personalization of LLMs, and machine learning. For more details, please check out my CV.

Publications

Weightless Fine-Tuning: Personalizing LLMs via Logit-Space Transport (COLM 2026)

Bohan Zhang, Anqi Ni, Yixin Wang, Paramveer Dhillon

Develop a training-free method for personalizing LLMs, approximating supervised fine-tuning by transporting supervised residuals in logit space during decoding.

Agentic Economic Modeling (ACL 2026 Industry Track)

Bohan Zhang, Jiaxuan Li, Ali Hortaçsu, Xiaoyang Ye, Victor Chernozhukov, Anqi Ni, Edward Huang

Introduce Agentic Economic Modeling (AEM), a framework that aligns synthetic choices generated by LLMs with small-sample human evidence for reliable econometric inference, and is validated in large-scale human experiments

Causal Inference for Human-Language Model Collaboration (NAACL 2024)

Bohan Zhang, Yixin Wang, Paramveer Dhillon

Introduce the Incremental Stylistic Effect and the CausalCollab algorithm to measure and leverage the causal impact of human editing styles in human-language model collaboration.

ExAnte: A Benchmark for Ex-Ante Inference in Large Language Models (EACL 2026)

Yachuan Liu, Xiaochun Wei, Lin Shi, Xinnuo Li, Bohan Zhang , Paramveer Dhillon, Qiaozhu Mei

We proposes ExAnte, a benchmark for evaluating ex-ante (time-constrained) reasoning in large language models, and shows that current models often leak future information despite explicit temporal restrictions.

Exploring Group-Level Signals for Robust Many-Domain Generalization (PAKDD 2025)

Bohan Zhang, Yachuan Liu, Qiaozhu Mei, Paramveer Dhillon

Propose a group-wise reweighting strategy using features like label entropy, representation statistics, and gradient properties to improve worst-group and tail performance for many-domain generalization.

More publications in CV →

Working Experience

Applied Scientist Intern

Mar 2025 - Oct 2025, May 2026 - Now

Stores Economics and Science (SEAS), Amazon

Developing a reasoning-distillation and GRPO fine-tuning pipeline that trains scalable reasoning LLMs to estimate how customer demand redistributes under catalog changes.

Designed LLM-based framework aligning LLM-generated synthetic responses with small-scale human data for reliable economic inference.

Machine Learning Software Engineering Intern

May 2021 - Aug 2021

Initium.AI

Deploy action sentence detection model for real-world applications with limited labels.

Student Instructor Assistant

Jan 2019 - May 2020

The Ohio State University

Education

Ph.D. in Information Science

Aug 2022 - Present

University of Michigan

M.S. in Computer Science and Engineering

Aug 2020 - Apr 2022

University of Michigan

B.S. in Computer Science

Aug 2017 - May 2020

The Ohio State University

Minor in Linguistics

Computer Science and Technology

Aug 2015 - Jul 2017

Xiamen University

Teaching

SI 671: Data Mining, University of Michigan

SI 315: Models of Social Information Processing, University of Michigan

CSE 2421: Introduction to Low-level Programming and Computer Organization, Ohio State University

Service

Regular Reviewer for *CL, WWW, KDD

Contact

🏀 🏊 🀄️ 🚣