About

This is Deqing Fu (傅德卿) and I'm a third-year Ph.D. student in Computer Science at the University of Southern California (USC). My main research interests are deep learning theory, natural language processing, and the interpretability of AI systems. I'm (co-)advised by Prof. Vatsal Sharan of USC Theory Group and Prof. Robin Jia of Allegro Lab within USC NLP Group, and I'm working closely with Prof. Mahdi Soltanolkotabi and Prof. Shang-Hua Teng.

News

Research

You can find a full picture of my research on the publications page or my Google Scholar page. They belong to the following categories:

Algorithmic Perspectives on Large Language Models

How to think about LLMs and transformers architectures from a theoretical and algorithmic perspective?

• How transformer implements its capability of in-context learning? Is it really doing gradient descent in-context? (NeurIPS 2024)
• How pretrained LLMs compute simple arithmetic tasks? Memorization or Mechanisms? (NeurIPS 2024, arXiv 2025)
• Is there a unified notion to distinguish Transformers from other neural architectures? (ICLR 2025)
• Can we use classical decision theory to guide LLMs to make decisions under uncertainty? (ICLR 2025 Spotlight)

Synthetic Data and Multimodal Learning

How can we evaluate Multimodal LLMs/VLMs in a robust way? How could we improve them with synthetic data?

• Are Multimodal LLMs sensitive to the input modality of the same problem? (COLM 2024)
• Improving Text-to-Image models with VLM's feedback. (NAACL 2025)
• Evaluating Multimodal LLMs' hallucination rates by training a token-level reward model (TLDR). Improving Multimodal LLMs with TLDR at both training and inference time. (ICLR 2025)

Education

• University of Southern California, 2022 - Present
  • Ph.D. in Computer Science
• University of Chicago, 2016 - 2022
  • M.S. in Statistics
  • B.S. with Honors in Mathematics, Computer Science, and Statistics