Hamza Cherkaoui

Télécom SudParis

Paris, France

I develop methods to improve the diversity and distributional coverage of generative models, particularly diffusion and flow-matching models, with a focus on recovering missing modes, capturing rare events, and modeling heavy-tailed distributions across image, text, and tabular data.

I am a postdoctoral researcher at Télécom SudParis, where I work with Prof. Hélène Halconruy. My work combines mathematical analysis with reproducible PyTorch/SLURM research pipelines and empirical benchmarks.

Previously, I worked at Noah’s Ark Lab, Huawei Technologies in Paris, with Dr. Igor Colin. I completed my Ph.D. in the Parietal team at CEA, advised by Dr. Philippe Ciuciu, Dr. Claire Leroy, and Dr. Thomas Moreau.

For more details, see my CV, Google Scholar, and GitHub.

Research themes

Generative modeling. Developing diffusion and flow-matching methods for diverse and reliable generation across data modalities.
Distributional coverage and heavy tails. Improving rare-mode coverage, capturing rare events, and modeling extreme observations.
Transfer learning. Identifying beneficial sources, enabling effective knowledge transfer, and mitigating negative transfer.

selected papers

Poster, NeurIPS, 2026	Do Heavy Tails Help Diffusion? On the Subtle Trade-off Between Initialization and Training paper Analyzes the trade-off between heavy-tailed initialization, training error, tail coverage, and sample diversity in diffusion and flow-based models.
Poster, AISTATS, 2026	High-Dimensional Analysis of Bootstrap Ensemble Classifiers paper code Uses random matrix theory to analyze bootstrap LSSVM ensembles and derive practical choices for subset selection and regularization.
Poster, ICML, 2025	Adaptive Sample Sharing for Multi Agent Linear Bandits paper Studies data sharing in multi-agent linear bandits and formalizes the bias-uncertainty trade-off behind collaboration.
Preprint, 2026	When to Transfer: Adaptive Source Selection for Positive Transfer in Linear Models paper Introduces a data-driven sample-sharing rule for using auxiliary data while protecting against negative transfer in ridge regression.
Preprint, 2026	Radial Tail Repair: Post-Training Correction for Diffusion and Flow Models paper Repairs the heavy-tail behavior of a frozen diffusion model by preserving its learned directions and correcting only the radial tail.

news

Sep 24, 2026	Our paper Do Heavy Tails Help Diffusion? On the Subtle Trade-off Between Initialization and Training was accepted as a poster in the NeurIPS 2026!
May 14, 2026	Our work High-Dimensional Analysis of Bootstrap Ensemble Classifiers was presented as a poster at AISTATS 2026 in Tangier.