Fengzhe Zhang

I am a PhD student at the Gatsby Computational Neuroscience Unit, supervised by Prof. Arthur Gretton. My PhD is fully funded by Jump Trading.

My research interests include generative modelling, variational inference, and probabilistic machine learning. I have worked on few-step generative models and efficient Boltzmann generators.

I interned at Jump Trading from July to September 2026. Previously, I worked on diffusion-based molecular samplers as a research assistant at the University of Cambridge, supervised by Prof. José Miguel Hernández-Lobato. As an undergraduate, I worked with Dr. Yingzhen Li on variational autoencoders (VAEs).

I hold an MPhil in Machine Learning and Machine Intelligence from Cambridge (Distinction) and a BSc in Mathematics from Imperial College London (First Class Honours).

news

Oct 2026	Started my PhD at the Gatsby Computational Neuroscience Unit.
Jul 2026	Presented our TMLR paper on variance-tuned diffusion models as a poster at ICML 2026 in Seoul.
Oct 2025	Our paper on variance-tuned diffusion models for efficient Boltzmann sampling was published in Transactions on Machine Learning Research (TMLR) and received Journal-to-Conference certification.
Dec 01, 2024	First-author work on efficient and unbiased Boltzmann sampling appeared at ML for Physical Sciences at NeurIPS 2024.
Nov 2024	Started as a Research Assistant at the University of Cambridge, working on diffusion-based generative samplers for molecular energy functions.

selected publications

TMLR
Efficient and Unbiased Sampling from Boltzmann Distributions via Variance-Tuned Diffusion Models

Fengzhe Zhang, Laurence I. Midgley, and José Miguel Hernández-Lobato

Transactions on Machine Learning Research, Oct 2025

Journal-to-Conference certification; presented as a poster at ICML 2026.

Abs Bib HTML PDF Code

Variance-Tuned Diffusion Importance Sampling (VT-DIS) improves Boltzmann sampling by tuning the noise covariance of a pretrained diffusion model. Trajectory-based importance weights correct estimation bias without costly likelihood calculations. Experiments on synthetic and molecular systems demonstrate improved sampling efficiency.
@article{zhang2025efficient, title = {Efficient and Unbiased Sampling from {Boltzmann} Distributions via Variance-Tuned Diffusion Models}, author = {Zhang, Fengzhe and Midgley, Laurence I. and Hern{\'a}ndez-Lobato, Jos{\'e} Miguel}, journal = {Transactions on Machine Learning Research}, year = {2025}, month = oct, url = {https://openreview.net/forum?id=Jq2dcMCS5R}, note = {Journal-to-Conference certification; presented as a poster at ICML 2026.} }
ML4PS
Efficient and Unbiased Sampling of Boltzmann Distributions via Consistency Models

Fengzhe Zhang, J. He, L. Midgley, and 2 more authors

In Machine Learning for Physical Sciences Workshop at NeurIPS, 2024

Abs Bib PDF

This work studies efficient, unbiased sampling of Boltzmann distributions by combining consistency models with importance sampling, reducing the number of function evaluations while preserving effective sample size on synthetic and equivariant n-body systems.
@inproceedings{zhang2024efficient, title = {Efficient and Unbiased Sampling of Boltzmann Distributions via Consistency Models}, author = {Zhang, Fengzhe and He, J. and Midgley, L. and Antor{\'a}n, J. and Hern{\'a}ndez-Lobato, Jos{\'e} Miguel}, booktitle = {Machine Learning for Physical Sciences Workshop at NeurIPS}, year = {2024}, }