I’m Jiazheng Li, a Ph.D. student in Computer Science at Tsinghua University, advised by Jingzhao Zhang. I work on machine learning through experiments and theory, with a particular interest in how large language models learn.

My current interests center on optimization for LLM pre-training and reinforcement learning for post-training. I like starting with an observation, testing it carefully, and building a theoretical explanation that reflects what actually happens in training.

Research interests

  • LLM optimization and training dynamics: understanding the behavior of neural networks and designing training algorithms.
  • Reasoning and reinforcement learning: supervised fine-tuning, RL post-training, and multi-turn agents.
  • Scalable model merging: combining model capabilities through progressive layer-wise distillation.

Education

Period Education
September 2025 – present Tsinghua University, Ph.D. in Computer Science, College of Artificial Intelligence. Advisor: Jingzhao Zhang.
September 2021 – June 2025 Beijing Institute of Technology, B.S. in Artificial Intelligence, School of Computer Science. GPA: 3.7/4.0.

Research experience

ByteDance Seed · LLM pre-training

Research intern · November 2025 – present

My work includes Lare, a loss-reweighting method for multi-domain LLM pre-training, and ActGD, an activation-aware optimizer designed to use unified hyperparameters across model parameters. These projects connect my interests in practical training behavior and optimization algorithms.

Ant Group RL Lab · LLM post-training

Research intern · April – September 2025

I worked on improving reasoning in language models with reinforcement learning and explored multi-turn agent approaches. This experience informs my interest in the interaction between training objectives and reasoning behavior.

Selected publications

Publication titles and venue distinctions below follow my October 2026 CV.

  • QuestA: Expanding Reasoning Capacity in LLMs via Question Augmentation · ICLR 2026, poster. Jiazheng Li, Hongzhou Lin, Hong Lu, Kaiyue Wen, Zaiwen Yang, Jiaxuan Gao, Yi Wu, Jingzhao Zhang.
  • Data Mixing Can Induce Phase Transitions in Knowledge Acquisition · NeurIPS 2025, spotlight. Xinran Gu, Kaifeng Lyu, Jiazheng Li, Jingzhao Zhang.
  • Scalable Model Merging with Progressive Layer-wise Distillation · ICML 2025, poster. Jing Xu, Jiazheng Li, Jingzhao Zhang.
  • NeuralDB: Scaling Knowledge Editing in LLMs to 90,000 Facts with Neural KV Database · ICLR 2026, poster. Weizhi Fei, Hao Shi, Jing Xu, Jingchen Peng, Jiazheng Li, Jingzhao Zhang, Bo Bai, Wei Han, Zhenyuan Chen, Xueyan Niu.

Things I build

LucidGPT is a GPT-style pre-training project with simple launch commands, multiple optimizer modes, benchmark evaluation, and optimizer/activation monitors for examining training stability.

For What-s-up, a chat software project, I worked on backend communication, functional testing, and debugging. AI4Game has an open project notebook where research and development notes can grow over time.

Beyond research

I enjoy algorithmic problem solving. Selected awards include gold at the 2022 ICPC Asia Hefei Regional Contest and the 2023 CCPC Invitational Contest in Hunan, alongside ICPC and CCPC regional silver medals. I work with Python, C++, C, Java and SQL, and write with LaTeX and Typst.

Why keep this notebook?

Papers tell a focused story. This notebook makes room for the original question, the assumption that failed, and the observation that suggested a better direction. Ideas, negative results and development decisions belong here alongside successful work.

Explore the research notebook, project showcases and visualization examples, or follow new entries through the RSS feed. Visualization demonstrations use explicitly labeled synthetic data; project descriptions and this introduction are adapted from my October 2026 CV.