When does Chain-of-Thought Help: A Markovian Perspective
arXiv, 2026
Identifies transition alignment as a key determinant of when chain-of-thought prompting improves sample efficiency, and validates the theory with controlled synthetic benchmarks.
Zihan Wang
I am a second-year Ph.D. student in the Department of Mathematics at the Courant Institute of Mathematical Sciences, New York University, advised by Qi Lei. My current research interests center on the mechanisms of supervised fine-tuning (SFT) and reinforcement learning (RL) in post-training, and on using mechanistic understanding to improve these methods. I also study chain-of-thought and reasoning in language models, aiming to understand their mechanisms through both theory and experiments. In addition, I am interested in optimization for LLMs and deep learning, as well as AI for mathematics.
News
- Attending ICML 2026 in Seoul, South Korea.
- When does Chain-of-Thought Help: A Markovian Perspective was posted on arXiv.
- Data Reconstruction Attacks and Defenses: A Systematic Evaluation was accepted to AISTATS 2025.
- Implicit bias of SGD in L2-regularized linear DNNs was accepted as an ICLR 2024 spotlight.
- Reconstructing Training Data from Model Gradient, Provably was accepted to AISTATS 2023.
Selected Publications
*: equal contribution.
Data Reconstruction Attacks and Defenses: A Systematic Evaluation
AISTATS, 2025
Recasts data reconstruction as an inverse problem, derives matching reconstruction-error bounds for two-layer networks, and proposes a utility-matched evaluation protocol.
Implicit bias of SGD in L2-regularized linear DNNs: One-way jumps from high to low rank
ICLR, 2024 Spotlight
Shows how stochastic gradient dynamics can move from higher-rank minima to lower-rank minima in regularized deep linear networks, while reverse jumps have probability zero.
Reconstructing Training Data from Model Gradient, Provably
AISTATS, 2023
Proves that a single model-gradient query at random initialization can identify training samples under broad neural network assumptions, with an efficient tensor-decomposition reconstruction algorithm.
Talks
- Reconstructing Training Data from Model Gradient, Provably. Machine learning reading group, University of California Berkeley.
Education
- Ph.D. in Mathematics, New York University. Moses A. Greenfield Research Prize.
- M.S. in Mathematics, New York University. Glenn Y. Louie Endowed Fellowship.
- B.S. in Statistics, Peking University.
Experience
Service
- Reviewer for ICLR, ICML, NeurIPS, and AISTATS.
Teaching
- TA for Honors Analysis I, MATH-UA 328.
- Grader for Real Variables I, MATH-GA 2430.
- Grader for Introduction to Partial Differential Equations, MATH-GA 2490.
Miscellaneous
- I enjoy soccer and support Real Madrid.
- I like modern detective novels. Here is my personal tier list.
- I play bridge, and represented NYU to place third in the 2025 NABC Collegiate Bridge Bowl.