Jingzhe Shi

Hi! I'm Jingzhe Shi (史景喆), previously an undergraduate student majoring Computer Science and Technology at IIIS, Tsinghua University (a.k.a Yao Class, directed by the Turing Award Laureate Andrew Chi-Chih Yao).

Recently, I am privileged to have the opportunity to be advised by and learn from great professors. Recently, I am honored to be advised by and to be collaborating with Prof. Mengdi Wang and Prof. Sanfeng Wu at Princeton, Prof. James Zou at Stanford, and Prof. Hang Zhao at Tsinghua on projects about LLM agent, LLM reasoning, LLM4Science, Scaling Laws and explanations, etc.

I am lucky to have collaborated with Qinwei Ma, my high school & university classmate. Previously, I was fortunate to have collaborated with Professor Xiaolong Wang and Doctor Lei Li, currently a Post-doc at UW.

I enjoyed learning Physics and contributing to Physics Olympiads. I won a gold medal at IPhO 2021 (ranking 10th globally), and I was an invited online marker for IPhO 2022.

📬 Email | 📑 CV | 🎓 Google Scholar | 💻 Github

Interests

Research interests: machine learning, large language models, scaling laws and theoretical foundations of deep learning, and reasoning in artificial intelligence systems.

Selected Publications & Preprints (in time order)

(* for equal contribution, ^ for equal correspondence)

	Oblivionis: Intrinsic Entropy of Context Length Scaling in LLMs Jingzhe Shi, Qinwei Ma, Hongyi Liu, Hang Zhao^, Jeng-Neng Hwang, Lei Li^ ICLR 2026 Oral* Code / arXiv We explain context length scaling from Intrinsic Space perspective, with theoretical assumptions and deductions validated by experiments on Natural Language and Synthetic Dataset. We find that Intrinsic Entropy, a metric measured from middle-states of LLMs, shows linear relationship to next token prediction loss, which could potentially be an interesting phenomenon to explore, and potentially implies the successes of sparse representations.
	PRISM-Physics: Causal DAG-Based Process Evaluation for Physics Reasoning Wanjia Zhao, Qinwei Ma, Jingzhe Shi, Shirley Wu, Jiaqi Han, Yijia Xiao, Si-Yuan Chen, Xiao Luo, Ludwig Schmidt, James Zou ICLR 2026* Project Page / arXiv We proposed Prism-Physics, a Physics Olympiad Problem dataset with DAG-based scoring policy and rule-based equation equivalence comparison, to evaluate LLMs' physics reasoning abilities in a fine-grained process-score-based manner. Our scoring framework and method are designed to be adaptable to other datasets that feature math reasoning process.
	Physics Supernova: AI Agent Matches Elite Gold Medalists at IPhO 2025 Jiahao Qiu, Jingzhe Shi, Xinzhe Juan, Zelin Zhao, Jiayi Geng, Shilong Liu, Hongru Wang, Sanfeng Wu, Mengdi Wang NeurIPS 2025 LLM Eval Workshop (Oral), NeurIPS 2025 LAW Workshop Code / arXiv We proposed an agent-based framework for solving International Physics Olympiad Problems. We conducted experiments on IPhO 2025 Theory Problems, validating the effectiveness of agent-based approaches.
	Scaling Law for Time Series Forecasting Jingzhe Shi, Qinwei Ma, Huan Ma, Lei Li NeurIPS 2024 Code / arXiv / OpenReview We proposed a theoretical framework for Scaling Law for Time Series Forecasting, taking into account look back horizon as well as dataset size and model size. We conducted experiments to validate our theory proposed and assumptions made. Our key theoretical and experimental findings were that optimal look back horizon does exist and it increases with dataset size, calling for a more fair comparison when proposing new time series forecasting models.
	CHOPS: CHat with custOmer Profile Systems for Customer Service with LLMs Jingzhe Shi, Jialuo Li, Qinwei Ma, Zaiwen Yang, Huan Ma, Lei Li COLM 2024 Code / arXiv / OpenReview We proposed CHOPS, an LLM agent designed to efficiently access user information, interact with existing systems, and provided accurate, safe responses by leveraging a combination of small and large LLMs. CHOPS demonstrated its potential to enhance or replace human customer service.
	Large Trajectory Models are Scalable Motion Predictors and Planners Qiao Sun, Shiduo Zhang, Danjiao Ma, Jingzhe Shi, Derun Li, Simian Luo, Yu Wang, Ningyi Xu, Guangzhi Cao, Hang Zhao arXiv 2023 Code / arXiv We leveraged successful backbones in NLP for trajectory prediction, demonstrating scalability on diverse datasets and achieving state-of-the-art performance on Nuplan dataset. I was responsible for the decoder part. I ustilized DDPM to generate trajectory in Key Point Space to capture multi-modal distribution of future trajectories.

Work Experience

Optiver
2024.07 - 2024.08
Trading Intern, Optiver Shanghai Office

Optiver is a leading global market-making firm headquartered in the Netherlands, with offices in financial centers worldwide.

Gained hands-on exposure to electronic market making, options pricing, and quantitative trading strategies.

Education Experience

	Tsinghua University 2021.09 - 2025.06 Bachelor of Engineering majoring in Computer Science at Yao Class, IIIS, Tsinghua GPA: 3.86/4.0 (overall), 3.90/4.0 (specialized courses).
	University of California, San Diego 2024.02 - 2024.06 Exchange Research Internship (J1 visa) Focus on research on time series and agent systems.

Honors and Awards

2023 & 2024: Technological Innovation Scholarship of Tsinghua University.

2024: Excellent Voluntary and Public Service Scholarship of Tsinghua University.

2021: First-Class Freshmen Scholarship of Tsinghua University.

2021: Gold Medal in the 51st International Physics Olympiad (IPhO 2021), ranking 10th globally among more than 360 contestants from approximately 70 countries and regions.

2020 - 2021: Member of the Chinese National Team for the International Physics Olympiad.

2020: Gold Medal in the Chinese Physics Olympiad (CPhO).

Language and Skills

Language: Chinese (Native), English (TOEFL: 113; R30,L28,S26,W29; test taken in Nov. 2024.), Japanese (daily dialogue).

Programming languages: Python, C/C++, etc.

Tools: Git, LaTeX, SQL, etc.

Service

Peer Reviewer: ICLR 2025 & 2026, COLM 2025 & 2026, NeurIPS 2025.

Invited Online Marker, 52nd International Physics Olympiad (IPhO 2022), held in Switzerland.

This homepage is designed based on Jon Barron's homepage and deployed on GitHub Pages. Last updated: May, 2026.
© 2026 Jingzhe Shi