Dario Shariatian

PhD researcher in generative modeling

Efficient generation · discrete diffusion · model systems

I am a PhD researcher at INRIA Paris, in the Sierra Project Team, led by Francis Bach. I have been advised by Umut Simsekli and Alain Durmus.

My research is in probabilistic generative modeling, especially continuous and discrete diffusion models. Recently, my focus has shifted toward efficient structured generation: hybrid autoregressive/diffusion methods, improved algorithms, inference-time acceleration with parallel generation or distillation, and robust evaluation loops for these systems.

Before that, I obtained an MSc in Mathematics from the University of Oxford and a BSc/MSc from Ecole Polytechnique.

More details can be found in my CV.

selected publications

SPIGM@ICML
Latent-Augmented Discrete Diffusion Models (LADD)

Dario Shariatian, Alain Durmus, Umut Simsekli, and 1 more author

In ICML 2026 Workshop on Structured Probabilistic Inference & Generative Modeling, May 2026

Abs arXiv Bib Code

We introduce a learnable auxiliary latent channel for discrete diffusion, enabling cross-token modeling and improving few-step generation.
@inproceedings{shariatian2026ladd, title = {Latent-Augmented Discrete Diffusion Models (LADD)}, author = {Shariatian, Dario and Durmus, Alain and Simsekli, Umut and Peluchetti, Stefano}, booktitle = {ICML 2026 Workshop on Structured Probabilistic Inference {\&} Generative Modeling}, year = {2026}, month = may, }
SPIGM@ICML
Uniform Diffusion Models Revisited: Leave-One-Out Denoiser and Absorbing State Reformulation

Samson Gourevitch, Yazid Janati, Dario Shariatian, and 4 more authors

In ICML 2026 Workshop on Structured Probabilistic Inference & Generative Modeling, May 2026

Abs arXiv Bib Code

We revisit uniform diffusion models, identify a leave-one-out denoising target, derive conversions between parameterizations, and introduce an absorbing-state reformulation with improved sampling behavior.
@inproceedings{gourevitch2026uniform, title = {Uniform Diffusion Models Revisited: Leave-One-Out Denoiser and Absorbing State Reformulation}, author = {Gourevitch, Samson and Janati, Yazid and Shariatian, Dario and Simsekli, Umut and Moulines, {\'E}ric and Xing, Eric P. and Durmus, Alain}, booktitle = {ICML 2026 Workshop on Structured Probabilistic Inference {\&} Generative Modeling}, year = {2026}, month = may, }
NeurIPS
Algorithm and Data Dependent Generalization Bounds for Score-Based Generative Models

Dario Shariatian, Benjamin Dupuis, Maxime Haddouche, and 2 more authors

In Advances in Neural Information Processing Systems, Sep 2025

Abs arXiv Bib Code

We establish generalization bounds for diffusion models that account jointly for optimization dynamics and data distribution, together with supporting empirical results.
@inproceedings{shariatian2025generalization, title = {Algorithm and Data Dependent Generalization Bounds for Score-Based Generative Models}, author = {Shariatian, Dario and Dupuis, Benjamin and Haddouche, Maxime and Durmus, Alain and Simsekli, Umut}, booktitle = {Advances in Neural Information Processing Systems}, year = {2025}, month = sep, }
ICML
Bit-Level Discrete Diffusion with Markov Probabilistic Models: An Improved Framework with Sharp Convergence Bounds under Minimal Assumptions

Dario Shariatian, Le-Tuyet-Nhi Pham, Antonio Ocello, and 2 more authors

In International Conference on Machine Learning, Feb 2025

Abs arXiv Bib Code

We develop a discrete diffusion method for bit data that outperforms MD4 and Discrete Flow Matching on binarized MNIST while using about 2.5x fewer network evaluations, with supporting theoretical analysis.
@inproceedings{shariatian2025dmpm, title = {Bit-Level Discrete Diffusion with Markov Probabilistic Models: An Improved Framework with Sharp Convergence Bounds under Minimal Assumptions}, author = {Shariatian, Dario and Pham, Le-Tuyet-Nhi and Ocello, Antonio and Conforti, Giovanni and Durmus, Alain}, booktitle = {International Conference on Machine Learning}, year = {2025}, month = feb, }

ICLR

Heavy-Tailed Diffusion with Denoising Levy Probabilistic Models

Dario Shariatian, Umut Simsekli, and Alain Durmus

In International Conference on Learning Representations, Jan 2025

Abs arXiv Bib Code Slides

@inproceedings{shariatian2025dlpm,
  title = {Heavy-Tailed Diffusion with Denoising Levy Probabilistic Models},
  author = {Shariatian, Dario and Simsekli, Umut and Durmus, Alain},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  month = jan,
}

NeurIPS
Piecewise Deterministic Generative Models

Andrea Bertazzi, Dario Shariatian, Umut Simsekli, and 2 more authors

In Advances in Neural Information Processing Systems, Jul 2024

Abs arXiv Bib Code

We introduce a class of generative models based on piecewise-deterministic Markov processes, combining deterministic motion with random jumps, and report promising empirical results.
@inproceedings{bertazzi2024pdmp, title = {Piecewise Deterministic Generative Models}, author = {Bertazzi, Andrea and Shariatian, Dario and Simsekli, Umut and Moulines, {\'E}ric and Durmus, Alain}, booktitle = {Advances in Neural Information Processing Systems}, year = {2024}, month = jul, }