avatar

Muzi Tao

Ph.D. Student
University of Southern California
muzitao (at) usc (dot) edu


About Me

I am a second-year Ph.D. student in Computer Science at the University of Southern California, advised by Prof. Xuezhe Ma. I received my Bachelor’s degree from the ACM Honors Class at Shanghai Jiao Tong University.

Previously, I was a visiting student researcher at New York University, fortunately advised by Prof. Saining Xie. Before that, I worked as a research intern at SJTU, advised by Prof. Li Niu.

My research interests lie broadly in multimodal generation and detailed visual processing capability in multimodality. Feel free to reach out if you are interested in my work or would like to connect.

Publications

  1. In Submission
    Lost in Translation: On the Idiosyncratic Gap between Image Captioning and Generation Models
    Muzi Tao, Chufan Shi, Huijuan Wang, Shengbang Tong, Xuezhe Ma
    In Submission. Coming soon
    We analyze idiosyncrasies in captioning models and their transfer to images using a simple attribution framework. While captions can be attributed to their source with near-perfect accuracy, these fingerprints weaken in images, revealing a cross-modal gap. By tracing stylistic cues in wording, color, and composition, we show how attribution can serve as a new lens for evaluating prompt following in text-to-image models.

  2. Tiny Papers, ICLR
    What Does a Visual Formal Analysis of the World's 500 Most Famous Paintings Tell Us About Multimodal LLMs?
    Muzi Tao, Saining Xie
    Tiny Papers, ICLR 2024. Invite to present, notable
    We present ArtQA, a benchmark that tests multimodal LLMs through the language of art. With questions on line, shape, color, and texture across 500 paintings, it offers a rigorous, expert-validated evaluation of visual understanding.
    PDF

Education

TA Experience

Current at the University of Southern California

Past at Shanghai Jiao Tong University


© 2025 Muzi Tao | Last updated: October 2025