Yu Li
George Washington University, Washington, D.C. 2025.9– Present
Wuhan University, Hongyi Honor College, China 2021.9-2025.5
Research Topics: Reinforcement Learning • Post-Training • Generative AI
I am currently a first-year Ph.D. student in the ECE Department at George Washington University, supervised by Prof. Tian Lan. I obtained my B.Eng. degree from the Hongyi Honor College of Wuhan University.
My research interests focus on Generative AI, Reinforcement Learning (RL), and Post-training methods for LLMs. I am always open to communication and collaboration.
- 📄 CV (PDF)
- 🧪 GitHub
- 🎓 Google Scholar
- ✉️ yul@gwu.edu
News
- [04/2025] I started my ECE PhD journey at GWU, supervised by Prof.Tian Lan.
- [03/2025] I joined the AGI Lab for a pre-graduation intern, focusing on generative models.
- [08/2024] I was awarded the Innova Excellence Scholarship (Top 3%).
Selected Projects
Unlocking Implicit Self-Reflection in Preference Optimization for LLM Alignment
Leveraging implicit preference information within preference pairs to establish a self-improvement mechanism, generalizing the theoretical foundation of existing preference optimization methods to enhance LLM alignment.
DPOSimPOPreference Learning
Aligning LLMs with Finite State Machine Logic for Multi-turn Verilog Code Generation
Enabling LLMs to learn state transition logic of finite state machines through structured alignment, constructing a multi-turn generation paradigm for Verilog code synthesis.
RLVRCode GenerationVerilog FSM
CRAFT-LORA: Content-Style Personalization via Rank-Constrained Adaptation
Enhancing content-style LoRA decomposition through rank-space constrained fine-tuning, and achieving personalized image generation via prompt mapping and asymmetric CFG for style-content LoRA fusion.
Generative AIPersonalizationLoRA
Prada: Black-Box LLM Adaptation with Private Data on Devices
· Paper
· Paper
Achieving efficient black-box LLM adaptation on edge device systems through probability differential methods while robustly preserving data privacy.
Black-Box LLMEdge AIPrivacy
Vision-Language Model-Guided Uncertainty-Aware Cross-Modal Sensor Fusion for Autonomous Vehicles
Our main work is to construct our VLM to improve the autonomous driving mode in abnormal weather conditions, focusing on improving the multi-modal fusion algorithm of Camera-Lidar branches.
Autonomous DrivingSensor FusionVision Language Model
DLoRA-TrOCR: Mixed Text Mode Optical Character Recognition Based On Transformer
· Paper · Code
TrOCR-based OCR with efficient PEFT for mixed text; practical pipeline and evaluation.
OCRLoRA
Dual branch SAM-Transformer Fusion Network for Accurate Breast Ultrasound Image Segmentation
We leveraged the rich semantic segmentation information of SAM and applied its fine-grained attention capability to the feature extraction module of Transformer, achieving SOTA IoU scores in ultrasound image segmentation.
Ultrasound SegmentationSAMTransformer
SfMDiffusion: Self-supervised Monocular Depth Estimation in Endoscopy Based on Diffusion Models
For endoscope medical scenarios, we use the diffusion model for depth estimation. We build a teacher model, set knowledge distillation, optical appearance and ddim losses, and introduce the teacher's discriminative prior, which significantly enhances the accuracy and confidence of the results.
Depth EstimationDiffusion ModelDistillation
Experiences
Mobile Intelligence Lab, George Washington University
Artificial General Intelligence Lab, Westlake University
Cyber-Physical Systems Lab, UC Irvine
Honors and Awards
- Innova International Exchange Scholarship, Wuhan University, 2024
- Innova Excellence Scholarship (Top 3%), Wuhan University, 2023, 2024
- Academic Excellence Scholarship (Top 5%), Hongyi Honor College, 2022, 2023, 2024
- First-Class Scholarship (Top 5%), Wuhan University, 2022, 2023, 2024
- Patent: Energy-saving calculation method, CN116085952.
Skills
- Languages: English (TOEFL 110), Chinese (Native), Japanese (N5)
- Programming: Python, C/C++, Matlab, Verilog
- Tools & Platforms: Ubuntu, Docker, Pytorch, Tensorflow, Git, Cadence
This website was stolen from my best friend CD.