Jingwei Li

Jingwei Li (李经纬) IIIS, Tsinghua University Email: ljw22 [at] mails.tsinghua.edu.cn Google Scholar

About Me

I am a final-year PhD student at the Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University, advised by Professor Jingzhao Zhang. Previously, I received my B.S. degree in Mathematics from the Department of Mathematical Sciences at Tsinghua University in 2022, and was selected for the Yau Mathematical Sciences Talent Program (Mathematics Xuetang Class).

Research Interests

My research lies at the intersection of theoretical and applied machine learning. On the theoretical side, I am interested in establishing provable guarantees for neural network generalization and control theory. On the applied side, my focus is on pre-training data for large language models (LLMs), such as optimizing data mixtures and exploring their effects on model performance.

Publications

α–β indicates alphabetical author order, * indicates equal contribution.

• Capacity-Aware Mixture Law Enables Efficient LLM Data Optimization [arxiv]
  Jingwei Li*, Xinran Gu* , Jingzhao Zhang
  International Conference on Machine Learning (ICML) 2026
  

• Finite Sample Analyses for Continuous-time Linear Systems: System Identification and Online Control [arxiv]
  Hongyi Zhou*, Jingwei Li* , Jingzhao Zhang
  Conference on Neural Information Processing Systems (NeurIPS) 2025
  

• Towards Black-box Membership Inference Attack for Diffusion Models [arxiv]
  Jingwei Li , Jing Dong, Tianxing He, Jingzhao Zhang
  International Conference on Machine Learning (ICML) 2025
  

• Understanding Nonlinear Implicit Bias via Region Counts in Input Space [arxiv]
  Jingwei Li* , Jing Xu*, Zifan Wang, Huishuai Zhang, Jingzhao Zhang
  International Conference on Machine Learning (ICML) 2025
  

• MutualNeRF: Improve the Performance of NeRF under Limited Samples with Mutual Information Theory [arxiv]
  Zifan Wang*, Jingwei Li* , Yitang Li, Yunze Liu
  International Conference on Uncertainty in Artificial Intelligence (UAI) 2025
  

• Online Control with Adversarial Disturbance for Continuous-time Linear Systems [arxiv]
  Jingwei Li , Jing Dong, Can Chang, Baoxiang Wang, Jingzhao Zhang
  Conference on Neural Information Processing Systems (NeurIPS) 2024
  

• Online Policy Optimization for Robust Markov Decision Process [arxiv]
  (α–β) Jing Dong, Jingwei Li , Baoxiang Wang, Jingzhao Zhang
  International Conference on Uncertainty in Artificial Intelligence (UAI) 2024
  

• Iteratively Learn Diverse Strategies with State Distance Information [arxiv]
  Wei Fu, Weihua Du* , Jingwei Li* , Sunli Chen, Jingzhao Zhang, Yi Wu
  Conference on Neural Information Processing Systems (NeurIPS) 2023
  

Preprints

• Scaling Domain Data Repetition in LLM Pretraining [arxiv]
  Jingwei Li, Xinran Gu, Rui Dai, Xintong Hao, Chengyin Xu, Yan Wu, Shuran Zheng, Jingzhao Zhang
  

Technical Report

• KIMI K2.5: Visual Agentic Intelligence [arxiv]
  Kimi Team et al.
  

Internship Experience

• Tencent (Hunyuan) 2026.10 - Present
  Algorithm Intern at Pre-training Team
  Focused on data mixture optimization, synthetic data generation, and pre-training evaluation for large language models (LLM).
• Jump Trading 2026.08 - 2026.09
  Quantitative Research Summer Intern
  Selected for the 2026 Quantitative Research Summer Internship Program to work on systematic trading strategy analysis.
• ByteDance (Seed) 2026.03 - 2026.07
  Algorithm Intern at Pre-training Team
  Focused on multi-epoch training for large language models (LLM) mid-training. Worked on developing efficient strategies for optimizing training epoch numbers for each dataset.
• Moonshot AI (KIMI) 2025.07 - 2026.02
  Algorithm Intern at Pre-training Team
  Focused on data mixture for large language models (LLM) mid-training. Worked on developing efficient strategies for optimizing data mixture. Contributed to the development of KIMI K2.5 pre-training phase.
• Microsoft Research Asia (MSRA) 2023.08 - 2023.12
  Research Intern at Machine Learning Theory Team
  Worked on deep learning theory, specifically on measures to describe implicit bias in neural networks. Successfully achieved a correlation of 0.98 between the generalization gap and the measure.