Deqing Fu

profile.jpg

This is Deqing Fu and I’m a last-year PhD candidate in Computer Science at the University of Southern California (USC). My main research interests are theoretical foundations of large language models and multimodal LLMs. 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. During my Ph.D. studies, I spent time at Google and Meta as a student researcher. Before USC, I completed my undergraduate degree in Mathematics (with honors) and my master’s in Statistics at the University of Chicago.

My research focuses on understanding large language models from algorithmic and theoretical perspectives, as well as developing practical methods in interpretability, synthetic data generation, and multimodal learning. You can find my publications on Google Scholar and my recent CV here.

Availability

I am on the job market in 2026–2027. Please reach out!

Research Highlights

Algorithmic Perspectives on Large Language Models
Interpretability and Alignment
Multimodal Models and Applications
  • Multimodal rewards for improving generation quality: token-level hallucination reduction (ICLR 2025) and Text-to-Image alignment (NAACL 2025)
  • Modality sensitivity in Multimodal LLMs (COLM 2024)
  • Large-scale dataset for visual reasoning with images (ICLR 2026)

News

all →
Jul 08, 2026 Convergent Evolution and Resa are accepted to COLM 2026!
Jun 30, 2026 I contributed to TabFM, a zero-shot foundation model for tabular data, released by Google Research!
Jun 29, 2026 New preprint: Value-Aware Stochastic KV Cache Eviction for Reasoning Models.
Jun 29, 2026 New blog post: Seeing Is Not Reasoning: How VLMs and Their Benchmarks Lean on Text.
Apr 30, 2026 Two papers (EPSVec and Transformers Learn Graph Connectivity) accepted to ICML 2026.
Apr 22, 2026 New preprint: Convergent Evolution: How Different Language Models Learn Similar Number Representations. See the website and blog post.
Jan 31, 2026 New preprint: EPSVec: Efficient and Private Synthetic Data Generation via Dataset Vectors.
Jan 26, 2026 Two papers (FoNE and Zebra-CoT) accepted to ICLR 2026.

Selected Publications

See full list or Google Scholar for all publications.

2026

  1. ICML
    graph.png
    Transformers Provably Learn Algorithmic Solutions for Graph Connectivity, But Only with the Right Data
    Qilin Ye*, Deqing Fu*, Robin Jia, and Vatsal Sharan
    In International Conference on Machine Learning (ICML), 2026
    *Equal Contribution
    📄 Paper PDF
  2. COLM
    convergent.png
    Convergent Evolution: How Different Language Models Learn Similar Number Representations
    Deqing Fu, Tianyi Zhou, Mikhail Belkin, Vatsal Sharan, and Robin Jia
    In Conference on Language Modeling (COLM), 2026
    📄 Paper Blog Website
  3. ICLR
    zebra-cot.png
    Zebra-CoT: A Dataset for Interleaved Vision Language Reasoning
    Ang Li*, Charles Wang*, Deqing Fu*, Kaiyu Yue*, Zikui Cai*, Wang Bill Zhu*, Ollie Liu* , Peng Guo*, Willie Neiswanger, Furong Huang, Tom Goldstein, and Micah Goldblum
    In International Conference on Learning Representations (ICLR), 2026
    *Equal Contribution
    📄 Paper Hugging Face Dataset
  4. ICLR
    fone.png
    FoNE: Precise Single-Token Number Embeddings via Fourier Features
    Tianyi Zhou, Deqing Fu, Mahdi Soltanolkotabi, Robin Jia, and Vatsal Sharan
    In International Conference on Learning Representations (ICLR), 2026
    📄 Paper Website
  5. ACL
    saegull.png
    Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models
    Woody Haosheng Gan*, Deqing Fu*, Julian Asilis*, Ollie Liu*, Dani Yogatama, Vatsal Sharan, Robin Jia, and Willie Neiswanger
    In Association of Computational Linguistics (ACL), 2026
    *Equal Contribution
    📄 Paper Blog

2025

  1. ICLR
    tldr.png
    TLDR: Token-Level Detective Reward Model for Large Vision Language Models
    Deqing Fu, Tong Xiao , Rui Wang, Wang Zhu, Pengchuan Zhang, Guan Pang, Robin Jia, and Lawrence Chen
    In International Conference on Learning Representations (ICLR), 2025
    📄 Paper
  2. ICLR
    sensitivity.png
    Transformers Learn Low Sensitivity Functions: Investigations and Implications
    Bhavya Vasudeva*, Deqing Fu*, Tianyi Zhou, Elliot Kau , You-Qi Huang, and Vatsal Sharan
    In International Conference on Learning Representations (ICLR), 2025
    *Equal Contribution
    📄 Paper
  3. ICLR
    dellma.png
    DeLLMa: Decision Making Under Uncertainty with Large Language Models
    Ollie Liu*, Deqing Fu*, Dani Yogatama, and Willie Neiswanger
    In International Conference on Learning Representations (ICLR), 2025
    Spotlight (Top 5.1%), *Equal Contribution
    📄 Paper GitHub Code Website
  4. NAACL
    dreamsync.png
    DreamSync: Aligning Text-to-Image Generation with Image Understanding Feedback
    Jiao Sun*, Deqing Fu*, Yushi Hu* , Su Wang, Royi Rassin, Da-Cheng Juan, Dana Alon, Charles Herrmann, Sjoerd Steenkiste, Ranjay Krishna, and Cyrus Rashtchian
    In Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2025
    *Equal Contribution
    📄 Paper

2024

  1. NeurIPS
    transformer-icl.png
    Transformers Learn to Achieve Second-Order Convergence Rates for In-Context Linear Regression
    Deqing Fu, Tian-Qi Chen, Robin Jia, and Vatsal Sharan
    In Conference on Neural Information Processing Systems (NeurIPS), 2024
    SoCalNLP Symposium 2023 Best Paper Award
    📄 Paper GitHub Code
  2. NeurIPS
    fourier.png
    Pre-trained Large Language Models Use Fourier Features to Compute Addition
    Tianyi Zhou, Deqing Fu, Vatsal Sharan, and Robin Jia
    In Conference on Neural Information Processing Systems (NeurIPS), 2024
    📄 Paper
  3. COLM
    isobench.png
    IsoBench: Benchmarking Multimodal Foundation Models on Isomorphic Representations
    Deqing Fu*, Ruohao Guo*, Ghazal Khalighinejad*, Ollie Liu*, Bhuwan Dhingra, Dani Yogatama, Robin Jia, and Willie Neiswanger
    In Conference on Language Modeling (COLM), 2024
    *Equal Contribution
    📄 Paper Website