Chenglei Yu 余成磊
Ph.D. student at Westlake University & Zhejiang University
I am a Ph.D. student (since 2024) in the joint program of Westlake University and Zhejiang University, advised by Prof. Tailin Wu in the AI for Scientific Simulation and Discovery Lab.
My research develops generative models — flow matching, optimal transport, and stochastic bridges — for single-cell biology: predicting how cell populations respond to perturbations and reconstructing how they develop over time, as steps toward AI virtual cells. I have also contributed to work on scientific machine learning, stochastic optimal control, and LLM reasoning.
News
PACE is accepted to NeurIPS 2026.
BuildArena is accepted to ICML 2026.
scDFM and RealPDEBench are accepted to ICLR 2026; RealPDEBench is selected for an Oral (top 1.2%).
On the Guidance of Flow Matching is accepted to ICML 2025 as a Spotlight.
Started my Ph.D. at Westlake University & Zhejiang University.
Publications
* Equal contribution First or co-first author Full list on Google Scholar
Conference Papers
-
NeurIPS 2026
PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference
Recovers continuous developmental trajectories from unpaired, destructive single-cell snapshots by transporting cells along a learned metric that follows the data manifold.
@inproceedings{yu2026pace, title = {{PACE}: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference}, author = {Yu, Chenglei and Wang, Chuanrui and Liao, Bangyan and Wu, Tailin}, booktitle = {Advances in Neural Information Processing Systems}, year = {2026} } -
ICML 2026
BuildArena: A Physics-Aligned Interactive Benchmark of LLMs for Engineering Construction
An interactive benchmark in which LLM agents turn a one-sentence goal into rockets, bridges, and vehicles, then test them in a physics simulator.
@inproceedings{xia2026buildarena, title = {{BuildArena}: A Physics-Aligned Interactive Benchmark of {LLMs} for Engineering Construction}, author = {Xia, Tian and Gao, Tianrun and Deng, Wenhao and Wei, Long and Qian, Xiaowei and Yu, Chenglei and Wu, Tailin}, booktitle = {Forty-third International Conference on Machine Learning}, year = {2026} } -
ICLR 2026
scDFM: Distributional Flow Matching Model for Robust Single-Cell Perturbation Prediction
Predicts how a perturbation shifts the whole distribution of cell states with conditional flow matching, without assuming cell-level correspondences.
@inproceedings{yu2026scdfm, title = {{scDFM}: Distributional Flow Matching Model for Robust Single-Cell Perturbation Prediction}, author = {Yu, Chenglei and Wang, Chuanrui and Liao, Bangyan and Wu, Tailin}, booktitle = {The Fourteenth International Conference on Learning Representations}, year = {2026} } -
ICLR 2026 Oral
RealPDEBench: A Benchmark for Complex Physical Systems with Real-World Data
Pairs real-world measurements with simulations of complex physical systems to study how scientific ML models cross the sim-to-real gap.
@inproceedings{hu2026realpdebench, title = {{RealPDEBench}: A Benchmark for Complex Physical Systems with Real-World Data}, author = {Hu, Peiyan and Feng, Haodong and Liu, Hongyuan and Yan, Tongtong and Deng, Wenhao and Gao, Tianrun and Zheng, Rong and Zheng, Haoren and Yu, Chenglei and Wang, Chuanrui and Li, Kaiwen and Ma, Zhi-Ming and Zhou, Dezhi and Lu, Xingcai and Fan, Dixia and Wu, Tailin}, booktitle = {The Fourteenth International Conference on Learning Representations}, year = {2026} } -
ICML 2025 Spotlight
On the Guidance of Flow Matching
A unified framework for energy guidance in general flow matching; it yields new guidance methods and recovers classical ones as special cases.
@inproceedings{feng2025guidance, title = {On the Guidance of Flow Matching}, author = {Feng, Ruiqi and Yu, Chenglei and Deng, Wenhao and Hu, Peiyan and Wu, Tailin}, booktitle = {Forty-second International Conference on Machine Learning}, year = {2025} }
Preprints
-
arXiv 2026
Path Integral Value Matching for Linear Quadratic Stochastic Optimal Control
A value-based solver for stochastic optimal control: a recursive path-integral value function learned by temporal-difference matching, without full-trajectory simulation.
@article{liao2026pivm, title = {Path Integral Value Matching for Linear Quadratic Stochastic Optimal Control}, author = {Liao, Bangyan and Yu, Chenglei and Yang, Yuchen and Wang, Chuanrui and Song, Zhisheng and Liu, Peidong and Wu, Tailin}, journal = {arXiv preprint arXiv:2608.10777}, year = {2026} } -
arXiv 2025
Unlocking Reasoning Capabilities in LLMs via Reinforcement Learning Exploration
RAPO swaps the mode-seeking reverse-KL penalty in RL with verifiable rewards for a forward-KL, reward-aware objective, letting LLMs explore beyond the base model.
@article{deng2025rapo, title = {Unlocking Reasoning Capabilities in {LLMs} via Reinforcement Learning Exploration}, author = {Deng, Wenhao and Wei, Long and Yu, Chenglei and Wu, Tailin}, journal = {arXiv preprint arXiv:2510.03865}, year = {2025} }
Experience
-
Westlake University · Ph.D. Student
Department of Artificial Intelligence, School of Engineering · AI for Scientific Simulation and Discovery Lab, advised by Prof. Tailin Wu -
Zhejiang University · Ph.D. Student
Joint Ph.D. program with Westlake University