Deqing Fu

This is Deqing Fu and I’m a final-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.

Research Highlights

Algorithmic Perspectives on Large Language Models

Can Transformers learn algorithms simply from data? (NeurIPS 2024, ICML 2026)
Arithmetic in pretrained LLMs: memorization vs. mechanisms? (NeurIPS 2024, ICLR 2026, COLM 2026)
What distinguishes Transformers from other architectures? (ICLR 2025)

Interpretability and Alignment

Decision theory for LLM reasoning under uncertainty (ICLR 2025 Spotlight, EMNLP 2026)
Steering vectors for improved visual understanding (ACL 2026), and for efficient and privacy-preserving synthetic data generation (ICML 2026)
Mechanistic interpretability via SAEs and transcoders (COLM 2026, Tech Report)

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 news →

Aug 29, 2026	Our paper Are LLM Decisions Faithful to Verbal Confidence? was accepted to EMNLP 2026 Main Conference!
Jul 21, 2026	New paper TOPL: Token-Level Off-Policy Learning for Faithful Generation Under Distribution Shift, accepted to COLM 2026!
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.

Selected Publications

All publications →

2026

ICML

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
COLM

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
ICLR

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 Dataset
ICLR

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
ACL

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

ICLR

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