VT Vinh Tong Scholar
Portrait of Vinh Tong

PhD candidate · Stuttgart, Germany

Generative models
for the physical world.

I develop efficient and symmetry-aware diffusion and flow-based models, with applications in molecular modelling and scientific discovery.

View selected work Email GitHub LinkedIn

Max Planck Institute for Intelligent Systems

University of Stuttgart

Thesis submission expected February 2027

Research statement

My work asks how generative models can become faster, respect physical symmetries, and transfer from large-scale generation to consequential scientific problems.

I began my PhD studying structured representations and now focus on diffusion models, flow matching, equivariance, and efficient inference across images, molecules, crystals, point clouds, proteins, and chemical reactions.

Selected work

Research across efficiency, symmetry, and science.

All publications
Molecular structure emerging over the SymDrift generation trajectory, from a single point to a full conformer.

01

2026 · Co-first author · Under review

SymDrift

One-shot generative modelling under symmetries using optimal alignment and invariant embeddings, evaluated on molecular conformers and transition states.

One-step generationEquivarianceMolecules
Potential-energy surface showing a transition state refined toward the true saddle point.

02

Journal of Chemical Information and Modeling · 2026

Adaptive Transition-State Refinement

Learned equilibrium flows for refining low-fidelity transition-state structures, increasing successful localisation by 41% and accelerating high-level quantum optimisation threefold.

Flow matchingComputational chemistry
A generated protein structure overlaid on its reference, beside generated peptide conformers.

03

NeurIPS 2025 · Equal contribution

Orbit Diffusion

Rao–Blackwellized gradient estimators for lower-variance, single-pass training of symmetry-aware diffusion models for molecules, crystals, and proteins.

Code ↗

DiffusionGradient estimationScientific ML
Paired image samples comparing a baseline sampler against LD3 at the same step budget.

04

ICLR 2025 · Oral Presentation

Learning to Discretize Denoising Diffusion ODEs

Learned, sampler-specific time discretizations that improve few-step generation from pre-trained diffusion models without retraining the base model.

Code ↗

Efficient inferenceDiffusion ODEs
A denoising progression grid beside a generated sample image.

05

ICML 2026 · Black Forest Labs

Self-Flow

Self-supervised flow matching for scalable multi-modal synthesis. I co-developed the project and investigated semantic information emerging in its latent embedding layers.

Code ↗

Flow matchingRepresentation learningMultimodal

Experience

Research in academia and at production scale.

I enjoy moving between mathematical ideas, careful experiments, and implementation at scale.

2022 — now

PhD Researcher

MPI for Intelligent Systems & University of Stuttgart

Efficient and equivariant generative modelling for scientific applications. Advised by Mathias Niepert.

2025 — 2026

Applied ML Research Intern

Black Forest Labs

Developed an improved low-step schedule for Flux 2 and co-developed Self-Flow.

2020 — 2022

Research Resident

VinAI Research

Graph neural networks for knowledge-graph completion and multilingual alignment.

Academic contributions

01

Research communication

Oral presentation of LD3 at ICLR 2025.

02

Mentoring & teaching

Mentored one Bachelor's and two Master's theses; TA for Introduction to AI and Reinforcement Learning.

03

Peer review

Reviewer for ICLR, ICML, NeurIPS, EMNLP, ECCV, and WACV.

Contact

Let’s build generative AI for meaningful science.

I welcome conversations about research collaborations and Research Scientist or Applied Research opportunities in scientific discovery and health.