Four themes spanning trustworthy AI, causal inference, and the foundations of deep learning.
My research spans the gap between theory and deployed AI systems. On one end, rigorous statistical foundations for deep learning. On the other, concrete tools for measuring and enforcing fairness in large-scale online platforms and generative models. Across all themes, evaluation methodology is the connective thread: how we measure properties of AI systems shapes what we build, how we audit it, and ultimately how we regulate it.
Generative models are routinely audited for bias, yet the verdicts rarely agree across papers or translate into actionable deployment decisions. The core problem is not the models themselves: it is the evaluation. Prompt family, decoding temperature, refusal handling, and scoring rubric each shift the verdict substantially, yet current practice leaves these degrees of freedom implicit.
Our research quantifies how much evaluation choices drive fairness conclusions relative to model properties, diagnoses the recurring failure modes (prompt sensitivity, sampling instability, counterfactual inconsistency, intersectional blind spots), and proposes Fairness Cards as a standardized reporting artifact. A Fairness Card pins down the audit surface, making findings reproducible, comparable, and actionable for the first time.
As illustrative evidence, we conducted a controlled audit of Qwen2.5-7B-Instruct across 3,200 generations on a factorial grid of demographic slices, prompt families, decoding regimes, and seeds. The headline finding: the same model receives opposite fairness verdicts depending solely on which prompt family is used to evaluate it.
Bias in machine learning manifests most concretely in the systems people interact with daily: job recommendation platforms, advertising auctions, credit scoring, and content feeds. These systems operate on tabular data at industrial scale, with sensitive attributes (gender, ethnicity) often unavailable at training or inference time.
Our work addresses this at two levels. FairJob provides the first large-scale real-world dataset for fairness in job advertising, with a pragmatic method for deriving protected attribute proxies under privacy constraints. On the algorithmic side, we study the theoretical connections between statistical independence measures (such as Cramér-von Mises, HSIC, and distance correlation) and standard fairness criteria (demographic parity, equalized odds), and use these connections to design efficient, differentiable regularizers that give practitioners direct control over the fairness-utility trade-off.
Personalization at industrial scale is fundamentally a causal problem: we want to know the effect of an action (a bid, a recommendation, a discount) on an individual, then allocate actions to maximize aggregate value under a budget constraint. Standard supervised learning cannot answer this: it requires uplift modeling and causal treatment effect estimation.
Our work develops two complementary approaches. CUVET exploits the economic structure of advertising (specifically the diminishing-returns relationship between spend and value) to derive a closed-form optimal policy via recursive partitioning that equalizes the marginal cost per unit value across user cohorts. For discrete allocation settings, we study how to maximize the success probability of a policy allocation, giving operators a principled way to reason about risk in stochastic environments.
What distribution does a deep network implicitly place on its outputs before seeing any data? With Gaussian priors on weights, the output prior becomes heavier-tailed as depth increases, following a generalized Weibull-tail pattern. Understanding this behavior is foundational for prior design, calibration, and the theoretical analysis of deep learning.
During a PhD at Inria Grenoble, we established the connection between BNN priors and Weibull-tail distributions, introduced the Sub-Weibull class of distributions as a unifying framework generalizing sub-Gaussian and sub-exponential properties, and characterized how units in deep networks become dependent as depth grows. This line of work is collected and extended in a review of Bayesian neural networks published in Statistical Science.
Two threads are shaping our current work. The first extends the fairness evaluation agenda to generative image models: we are studying bias in diffusion models and how standard evaluation protocols fail in the visual generation setting in the same way they fail for text.
The second thread moves toward trustworthiness and safety in large language models more broadly, working with collaborators at the intersection of alignment, robustness, and evaluation methodology. As models shift from single-turn generation to agentic systems capable of autonomous multi-step reasoning and tool use, the evaluation challenges multiply: the space of possible behaviors grows combinatorially, failure modes are harder to anticipate, and the gap between offline benchmarks and deployed behavior widens. Understanding how to audit, constrain, and certify agentic systems is a pressing open problem.
The longer-term goal is to build evaluation into a first-class research area with its own principled methods, failure taxonomies, and reproducibility standards, grounded in current industry practices and directly informing emerging AI regulations.