About Me
I am currently a first-year Ph.D. student at the Gaoling School of Artificial Intelligence, Renmin University of China. I am a member of the Intelligent Information Retrieval Lab (IIR), supervised by Prof. Jun Xu and Assoc. Prof. Xiao Zhang. My research interests include Large Language Models, Reinforcement Learning, Information Retrieval, and Recommender Systems. I am always open to collaboration and discussions in these areas. If you are interested, please feel free to reach out!
Selected Awards
- 🥇 Gold Medal, ACM-ICPC Asia Regional Contest, Nanjing, 2024
- 🥈 Silver Medal, China Collegiate Programming Contest (CCPC), Zhengzhou, 2024
- 🥉 Bronze Medal, ACM-ICPC Asia Regional Contest, Hangzhou, 2024
- 🌍 Top 0.36% (32 / 8785) Worldwide, IEEE Xtreme 18 Programming Competition — 2nd in China, 2024
Publications
Efficient Activation Optimization for Large-Scale Model Training
Advisor: Prof. Yuan Kun, Peking University Duration: Mar 2025 – May 2025
- Focused on activation compression algorithms by analyzing the structured sparsity of activation tensors, and developed a dynamic quantization-based mixed-granularity compression scheme.
- Extended the COAT quantization framework to INT8 and implemented full-process verification on OLMo, achieving performance nearly lossless compared to the BF16 baseline.
- Explored mixed-granularity activation quantization, achieving 1.6× acceleration and 40.8% activation memory reduction on a single-layer Transformer, and 1.68× end-to-end training speedup with 35.1% peak memory reduction.
Multimodal Large Model Reasoning for Visual Puzzle Understanding
Institution: Shanghai Innovation Institute Duration: May 2025
- Addressed reasoning limitations of large models on visual grid puzzles by designing an integrated approach combining full-parameter fine-tuning, Chain-of-Thought (CoT) distillation, multi-source data fusion, and diversified prompt engineering.
- Designed a CoT distillation pipeline using Qwen2.5-VL-72B to generate reasoning process samples for data augmentation, applying a progressive fusion strategy to boost Qwen2.5-VL-7B-Instruct’s cross-task accuracy from 52% to 72%.
- Introduced VGRP-Bench and designed a perturbation-based data augmentation strategy, raising accuracy on high-difficulty tasks (Aquarium) from 79% to 92%.
- Achieved 4th place (Top 0.7%) in the Multimodal Reasoning Puzzle Challenge.
Honors and Scholarships
- 🌟 National Scholarship (Dec 2024, Dec 2023)
- 💡 CCF Outstanding Undergraduate Program (Oct 2025)
- ✨ First-Class Scholarship, Xidian University (Dec 2022)
Education
- 2026.09 – Present, Ph.D. Candidate, Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China
- 2022.09 – 2026.06, B.Eng., School of Computer Science, Xidian University, Xi’an, China
Internships
- 2025.10 – Present, Tencent WeChat Group, China
Invited Talks
- Invited to give a talk at the Programming Contest Seminar of Xidian University — news
Other
In my spare time, I enjoy participating in Programming Contests and playing chess. I have been playing chess since the age of four and hold the title of Master of the Chinese Chess Association. I was the Youth Champion of Anhui Province in 2012, 2013, and 2014.
