Publications — Mariia Vladimirova

2026

Fairness Failure in Generative Models is an Evaluation Problem

Mariia Vladimirova, Jean-Yves Franceschi, Thibaut Issenhuth

ICMLPosition Paper2026

Project page HAL Code & data

This position paper argues that fairness failures in generative models are ultimately an evaluation problem: fairness findings are rarely comparable across papers or actionable for deployment decisions. We diagnose recurring empirical and conceptual failure modes and propose Fairness Cards as a minimal reporting artifact that makes evaluation choices explicit (covering prompt protocols, decoding settings, refusal handling, and scorer details), enabling reproducibility, comparability, and accountability.

@inproceedings{vladimirova2026fairness, title={Position: Fairness Failure in Generative Models is an Evaluation Problem}, author={Vladimirova, Mariia and Franceschi, Jean-Yves and Issenhuth, Thibaut}, booktitle={International Conference on Machine Learning}, year={2026} }
A primer on Bayesian neural networks: review and debates

Julyan Arbel, Konstantinos Pitas, Mariia Vladimirova, Vincent Fortuin

Statistical Science2026 (accepted 2025)

PDF

A review of Bayesian neural networks covering their theoretical foundations, approximate inference methods, and practical applications, alongside a discussion of open debates in the field regarding scalability, prior choice, and calibration.

@article{arbel2026primer, title={A Primer on {Bayesian} Neural Networks: Review and Debates}, author={Arbel, Julyan and Pitas, Konstantinos and Vladimirova, Mariia and Fortuin, Vincent}, journal={Statistical Science}, volume={41}, number={2}, pages={316--353}, year={2026}, doi={10.1214/24-STS969} }

2025

Fairness in Generative AI is Understudied, Underachieved, Undervalued

Mariia Vladimirova, Jean-Yves Franceschi, Thibaut Issenhuth

PreprintHAL, 2025

HAL

A position paper arguing that fairness in generative AI is systematically understudied due to evaluation gaps, underachieved due to the complexity of generative systems, and undervalued in deployment contexts. We identify structural barriers and propose directions for progress.

@misc{vladimirova2025fairnessgenai, title={Fairness in Generative {AI} is Understudied, Underachieved, Undervalued}, author={Vladimirova, Mariia and Franceschi, Jean-Yves and Issenhuth, Thibaut}, year={2025}, howpublished={HAL preprint hal-05318171} }
Fairness Through Independence via Cramér-von Mises Regularization

Albert Gimó Contreras, Mariia Vladimirova, Olga Petrova, Federico Pavone, Reda Chhaibi

NeurIPSWorkshopReliable ML Workshop, 2025

We introduce a CvM-based regularizer that promotes statistical independence between model outputs and sensitive attributes, with an efficient differentiable sample-based implementation. We provide theoretical analysis connecting CvM to demographic parity and equalized odds, and demonstrate competitive utility-fairness trade-offs on tabular benchmarks.

@inproceedings{contreras2025cvm, title={Fairness Through Independence via {Cram\'{e}r-von Mises} Regularization}, author={Contreras, Albert Gim{\'o} and Vladimirova, Mariia and Petrova, Olga and Pavone, Federico and Chhaibi, Reda}, booktitle={NeurIPS Workshop on Reliable Machine Learning}, year={2025} }
CUVET: A Partitioning Approach for Continuous Treatment Assignment At Scale

Artem Betlei, Mariia Vladimirova, Victor Girou, Thibaud Rahier

NeurIPSWorkshopNeurIPS CauScien Workshop, 2025

Dataset

CUVET (Cost per Unit Value Equalization Tree) is a novel method for continuous treatment assignment that recursively partitions the user space to equalize marginal cost per unit value across cohorts. The method yields a closed-form cost-aware policy and scales to large digital advertising datasets. We also introduce the CUVET-policy benchmark, an open-source dataset derived from real-world A/B experiments.

@inproceedings{betlei2025cuvet, title={{CUVET}: A Partitioning Approach for Continuous Treatment Assignment At Scale}, author={Betlei, Artem and Vladimirova, Mariia and Girou, Victor and Rahier, Thibaud}, booktitle={NeurIPS Workshop on Uncovering Causality in Science}, year={2025} }

2024

FairJob: A Real-World Dataset for Fairness in Online Systems

Mariia Vladimirova, Eustache Diemert, Federico Pavone

NeurIPSDatasets & Benchmarks, 2024

PDF Dataset Blog post

We present FairJob, a real-world dataset for studying fairness in job recommendation within online advertising. A key contribution is a method for obtaining proxy estimates of protected attributes not directly observable, while preserving a realistic benchmark. Experiments demonstrate potential improvements in fairness and the associated trade-offs with utility when applying bias mitigation strategies.

@inproceedings{vladimirova2024fairjob, title={{FairJob}: A Real-World Dataset for Fairness in Online Systems}, author={Vladimirova, Mariia and Diemert, Eustache and Pavone, Federico}, booktitle={Advances in Neural Information Processing Systems, Datasets and Benchmarks Track}, year={2024} }
Maximizing the Success Probability of Policy Allocations in Online Systems

Artem Betlei, Mariia Vladimirova, Mehdi Sebbar, Nicolas Urien, Thibaud Rahier, Benjamin Heymann

AAAI2024

Paper

We study the problem of maximizing the probability that a policy allocation achieves a target value in online systems with stochastic responses. We propose efficient algorithms with theoretical guarantees and demonstrate effectiveness on real-world advertising allocation problems.

@inproceedings{betlei2024maximizing, title={Maximizing the Success Probability of Policy Allocations in Online Systems}, author={Betlei, Artem and Vladimirova, Mariia and Sebbar, Mehdi and Urien, Nicolas and Rahier, Thibaud and Heymann, Benjamin}, booktitle={Proceedings of the AAAI Conference on Artificial Intelligence}, volume={38}, pages={11061--11068}, year={2024} }

Earlier

Dependence between Bayesian Neural Network Units

Mariia Vladimirova, Julyan Arbel, Stéphane Girard

PreprintarXiv, 2021

arXiv

We study the dependence structure between units in Bayesian neural networks with Gaussian priors on the weights, showing that units become increasingly correlated as the number of layers grows, with implications for prior specification and approximate inference.

@misc{vladimirova2021dependence, title={Dependence between {Bayesian} Neural Network Units}, author={Vladimirova, Mariia and Arbel, Julyan and Girard, St{\'e}phane}, year={2021}, eprint={2111.14397}, archivePrefix={arXiv} }
Bayesian Neural Network Unit Priors and Generalized Weibull-Tail Property

Mariia Vladimirova, Julyan Arbel, Stéphane Girard

ACMLAsian Conference on Machine Learning, 2021

Scholar

We characterize the prior induced on Bayesian neural network outputs by Gaussian priors on weights, establishing connections to generalized Weibull-tail distributions. The results provide theoretical grounding for understanding depth-dependent heavy-tailedness in BNN priors.

@inproceedings{vladimirova2021bnn, title={{Bayesian} Neural Network Unit Priors and Generalized {Weibull}-Tail Property}, author={Vladimirova, Mariia and Arbel, Julyan and Girard, St{\'e}phane}, booktitle={Asian Conference on Machine Learning}, pages={1397--1412}, year={2021} }
Sub-Weibull Distributions: Generalizing Sub-Gaussian and Sub-Exponential Properties to Heavier-Tailed Distributions

Mariia Vladimirova, Stéphane Girard, Hien Nguyen, Julyan Arbel

StatWiley, 2020

Scholar

We introduce Sub-Weibull distributions as a unifying class generalizing sub-Gaussian and sub-exponential properties to heavier-tailed settings. We establish tail bounds, sum rules, and connections to statistical learning theory, providing a toolkit for modern high-dimensional analysis.

@article{vladimirova2020subweibull, title={Sub-{Weibull} Distributions: Generalizing Sub-{Gaussian} and Sub-Exponential Properties to Heavier-Tailed Distributions}, author={Vladimirova, Mariia and Girard, St{\'e}phane and Nguyen, Hien and Arbel, Julyan}, journal={Stat}, volume={9}, number={1}, pages={e318}, year={2020} }
Understanding Priors in Bayesian Neural Networks at the Unit Level

Mariia Vladimirova, Jakob Verbeek, Pablo Mesejo, Julyan Arbel

ICML2019

Scholar

We study the prior induced on deep Bayesian neural networks with Gaussian priors on weights, showing that output distributions become increasingly heavy-tailed with depth, following a Weibull-tail pattern. This provides theoretical grounding for prior design in practical BNNs and was among the first to formally characterize this depth-dependent behavior.

@inproceedings{vladimirova2019understanding, title={Understanding Priors in {Bayesian} Neural Networks at the Unit Level}, author={Vladimirova, Mariia and Verbeek, Jakob and Mesejo, Pablo and Arbel, Julyan}, booktitle={International Conference on Machine Learning}, pages={6458--6467}, year={2019} }