PhD in Causal Inference at Université Lorraine

Toward a Causal and Experimental Design Framework for AI Evaluation

This project aims to develop a causal and experimental design-based evaluation framework to assess the robustness of AI models. The goal is to move beyond static benchmarks by examining how models respond to controlled interventions.

We will build a framework grounded in causal invariance to identify mechanisms that remain stable across context shifts and distribution changes. By integrating design of experiments (DoE) methodologies, we will create targeted protocols to uncover model vulnerabilities with minimal interventions.

These methods will be consolidated within INESIA-DoE Lab, a platform dedicated to generating and executing experimental designs, accompanied by an open-source library combining causal tools and intervention strategies. Finally, a multi-level analysis will explore how local perturbations (in data or architecture) influence global behavior and contribute to error propagation.

Feel free to contact us for more information :

marianne.clausel@univ-lorraine.fr and emilie.devijver@univ-grenoble-alpes.fr