Deqing Fu
This is Deqing Fu and I’m a fourth-year Ph.D. candidate 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. 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.
Algorithmic Perspectives on Large Language Models
- Can Transformers learn algorithms simply from data? (NeurIPS 2024, arXiv 2025)
- Arithmetic in pretrained LLMs: memorization vs. mechanisms? (NeurIPS 2024, ICLR 2026)
- What distinguishes Transformers from other architectures? (ICLR 2025)
Interpretability and Alignment
- Decision theory for LLM reasoning under uncertainty (ICLR 2025 Spotlight, arXiv 2026)
- Steering vectors for improved visual understanding (arXiv 2025), and for efficient and privacy-preserving synthetic data generation (arXiv 2026)
- Mechanistic interpretability via SAEs and transcoders (arXiv 2025, 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
| 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. |
| Oct 22, 2025 | New preprint: Transformers Provably Learn Algorithmic Solutions for Graph Connectivity, But Only with the Right Data. |
| Jul 23, 2025 | Three new preprints (Multimodal Steering, Resa, and Zebra-CoT) |
| Jul 22, 2025 | Check out our new paper Zebra-CoT on interleaved text and visual reasoning! Dataset and model are available on Hugging Face 🤗. |
Selected Publications
See full list or Google Scholar for all publications.
2026
Dataset
2025
- arXiv
Transformers Provably Learn Algorithmic Solutions for Graph Connectivity, But Only with the Right DataIn arXiv, 2025*Equal Contribution - ICLR
TLDR: Token-Level Detective Reward Model for Large Vision Language ModelsIn International Conference on Learning Representations (ICLR), 2025 - ICLR
Transformers Learn Low Sensitivity Functions: Investigations and ImplicationsIn International Conference on Learning Representations (ICLR), 2025*Equal Contribution - NAACL
DreamSync: Aligning Text-to-Image Generation with Image Understanding FeedbackIn Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2025*Equal Contribution
Code
2024
- NeurIPS
Pre-trained Large Language Models Use Fourier Features to Compute AdditionIn Conference on Neural Information Processing Systems (NeurIPS), 2024
