Hi! I’m Yuyang Song, an incoming PhD student in Computer Science at University of Texas at Austin, where I feel fortunate to be advised by Prof. Yizhong Wang. I wrap up my B.S. in Computer Science at the Wu Yuzhang Honors College, Sichuan University, advised by Prof. Mingjie Tang. I also work closely with Prof. Jianguo Wang at Purdue University. Currently, I am a member of the foundation model group at IQuest AI Lab, Ubiquant.
I’m broadly interested in how data shapes a model as it flows through training, and how architecture and training should be designed around that flow, with efficiency as a guiding principle. Rather than treating the model, its data, and its training as separate problems, I see them as forces that continually shape one another, and I want to understand how to harness that interplay to build capable models more efficiently.
📝 Selected Publications
Large-Scale Terminal Agentic Trajectory Generation from Dockerized Environments | Code
Siwei Wu*, Yizhi Li*, Yuyang Song*, Wei Zahng, Riza Batista-Navarro, Xian yang, Mingjie Tang, Bryan Dai, Jian Yang, Chenghua Lin.
We study how large-scale, execution-grounded terminal trajectories reshape the agentic capabilities of code LLMs. Trained on this data, TerminalTraj-32B achieves state-of-the-art results among models under 100B parameters and matches models more than 10× larger, while showing markedly steeper test-time scaling—evidence that the right training data, not just scale, drives terminal agent capability.
IQuest-Coder-V1 Technical Report | Code
Jian Yang, Wei Zhang, Shawn Guo, Zhengmao Ye, Lin Jing, Shark Liu, Yizhi Li, Jiajun Wu, Cening Liu, X. Ma, Yuyang Song, et al.
We introduce IQuest-Coder-V1, a high-performance code LLM that achieves leading results across multiple public benchmarks. My contribution focused on post-training optimization, specifically Text-to-SQL reasoning and agentic terminal tasks, improving the model’s practical reasoning and reliability on real-world tasks.
QUITE: A Query Rewrite System Beyond Rules with LLM Agents | Code
Yuyang Song, Hanxu Yan, Jiale Lao, Yibo Wang, Yufei Li, Yuanchun Zhou, Jianguo Wang, Mingjie Tang
QUITE is a training-free and feedback-aware system based on LLM agents that rewrites SQL queries into semantically equivalent forms with significantly better performance, covering a broader range of query patterns and rewrite strategies compared to rule-based methods.
🎖 Honors and Awards
- 2026.05, Outstanding Undergraduate Graduate, Sichuan Province.
- 2023.11, First Prize, China Undergraduate Mathematical Contest in Modeling.
- 2023.10, National Scholarship, Ministry of Education, China (Top 1%).
- 2023.10, First-Class Scholarship, Sichuan University (Top 1%).
- 2023.08, First Prize, China Undergraduate Internet of Things Design Competition.
📖 Educations
2026.08 (incoming), Ph.D. in Computer Science, University of Texas at Austin.
2022.09 - 2026.06, B.S. in Computer Science, Wuyuzhang College (Honours Programme), Sichuan University.
💻 Internships
- 2025.09 - now, Research Internship, IQuest AI Lab, Ubiquant, China.
- 2024.11 - 2025.07, Research Assistant, Purdue University, USA.
- 2024.09 - 2025.07, Research Assistant, Sichuan University, China.