2025Activity reportProject-Team​​SODA

2.1.1 Application context: richer​​​‌ data in health and‌ social sciences

Opportunistic data‌​‌ accumulations, often observational, bare​​ great promises for social​​​‌ and health sciences. But‌ the data are too‌​‌ big and complex for​​ standard statistical methodologies in​​​‌ these sciences.

2.1.2 Related‌​‌ data-science challenges

7.1.1 Scikit-learn​​​‌

• Keywords:
  Clustering, Classification, Regression,‌ Machine learning
• Scientific Description:‌​‌
  Scikit-learn is a Python​​ module integrating classic machine​​​‌ learning algorithms in the‌ tightly-knit scientific Python world.‌​‌ It aims to provide​​ simple and efficient solutions​​​‌ to learning problems, accessible‌ to everybody and reusable‌​‌ in various contexts: machine-learning​​ as a versatile tool​​​‌ for science and engineering.‌
• Functional Description:

  Scikit-learn can‌​‌ be used as a​​ middleware for prediction tasks.​​​‌ For example, many web‌ startups adapt Scikitlearn to‌​‌ predict buying behavior of​​ users, provide product recommendations,​​​‌ detect trends or abusive‌ behavior (fraud, spam). Scikit-learn‌​‌ is used to extract​​ the structure of complex​​​‌ data (text, images) and‌ classify such data with‌​‌ techniques relevant to the​​ state of the art.​​​‌

  Easy to use, efficient‌ and accessible to non‌​‌ datascience experts, Scikit-learn is​​ an increasingly popular machine​​​‌ learning library in Python.‌ In a data exploration‌​‌ step, the user can​​​‌ enter a few lines​ on an interactive (but​‌ non-graphical) interface and immediately​​ sees the results of​​​‌ his request. Scikitlearn is​ a prediction engine .​‌ Scikit-learn is developed in​​ open source, and available​​​‌ under the BSD license.​

• URL:
  http://­scikit-learn.­org
• Publications:
  hal-00650905​‌, hal-00856511, hal-01093971​​
• Contact:
  Gael Varoquaux
• Participant:​​​‌
  10 anonymous participants
• Partners:​
  Axa, BNP Parisbas Cardif,​‌ Dataiku, Nvidia, Chanel, Probabl​​

7.1.2 joblib

• Keywords:
  Parallel​​​‌ computing, Cache
• Functional Description:​
  Facilitate parallel computing and​‌ caching in Python.
• URL:​​
  https://­joblib.­readthedocs.­io/­en/­latest/
• Contact:
  Thomas Moreau​​​‌
• Participant:
  an anonymous participant​
• Partner:
  Probabl

7.1.3 skrub​‌

• Keyword:
  Data analysis
• Functional​​ Description:
  Joins, aggregates, and​​​‌ vectorizes tables to enable​ statistical learning, including with​‌ badly formated entries
• URL:​​
  https://­skrub-data.­org
• Contact:
  Gael Varoquaux​​​‌
• Participant:
  2 anonymous participants​
• Partner:
  Probabl

TabICL: Table foundation models​​

A new wave of​​​‌ progress is pushing forward​ tabular learning. Recent models​‌ have been bringing better​​ performance across the board.​​​‌ A poster example is​ that of the TabPFN​‌ series of models, that​​ rely on pretrained transformers​​​‌ to bring excellent performance,​ originally in the few-shot​‌ settings, and in the​​ beginning of 2025, up​​​‌ to moderate tables with​ TabPFN2. This line of​‌ work has led to​​ spinning off a startup​​​‌ in Germany. However, the​ quadratic complexity of the​‌ transformers is a bottleneck.​​ With the TabICL model​​​‌ 7, we showed​ that a multi-stage architecture​‌ can build a pre-trained​​ in-context predictor where the​​​‌ separation of states decreases​ the quadratic cost. The​‌ model can be pretrained​​ on larger datasets, and​​​‌ thus results in the​ best performer in settings​‌ of larger tables. The​​ model is faster than​​​‌ alternatives, in particular when​ using a CPU rather​‌ than a GPU. In​​ addition, we released in​​​‌ open source all the​ code, including the pretraining;​‌ this has spurred much​​ downstream research for multiple​​​‌ applications and enhancements, such​ as privacy.

This result​‌ is very significant as​​ it pushes forward the​​​‌ agenda of foundation models​ for tables. It is​‌ giving birth to a​​ very active line of​​​‌ research. The paper has​ been cited 72 times​‌ in less than a​​ year.

Retrieve merge predict​​​‌

A full data-science pipeline​ must often assemble data​‌ across multiple source tables.​​ When the user is​​​‌ faced with a complex​ data lake, many tables​‌ and little explicit link​​ between them, it is​​​‌ difficult to find the​ best assembly for a​‌ given machine-learning task. This​​ problem requires not only​​​‌ finding which table must​ be joined in the​‌ main table of interest​​ –a table retrieval problem–,​​​‌ but also how to​ aggregate multiple records when​‌ tables are linked through​​ a many-to-one relation. While​​​‌ table retrieval is a​ classic problem of the​‌ data management literature, it​​ had been understudied in​​​‌ the case of supervised​ machine learning. We assembled​‌ a systematic –and open–​​ benchmark with data lakes​​​‌ and supervised-learning tasks 2​. We found that​‌ supervised learning does change​​ the picture compared to​​ classic table-retrieval settings in​​​‌ that for a fixed‌ compute budget, it is‌​‌ worth avoiding fancy retrieval​​ methods, which can be​​​‌ very computationally costly, and‌ rather using better supervised‌​‌ learning methods, which can​​ be comparatively less expensive​​​‌ while being able to‌ extract the relevant information‌​‌ from a noisy retrieval.​​

TabArena

The progress in​​​‌ tabular learning—using machine learning‌ to predict from rows‌​‌ of a table—has been​​ driven by empirical studies​​​‌ over the last few‌ years. We have contributed‌​‌ to building TabArena 4​​, a living benchmark​​​‌ for machine learning on‌ tabular data. TabArena contains‌​‌ 51 datasets that are​​ carefully curated to represent​​​‌ real-world tabular learning tasks,‌ avoiding pitfalls such as‌​‌ duplicated datasets with different​​ names, data leakage, inappropriate​​​‌ train-test splits, datasets inappropriate‌ for tabular learning methods,‌​‌ and so on. The​​ first version of the​​​‌ benchmark evaluates 16 tabular‌ learning methods, including recent‌​‌ models and 3 table​​ foundation models. TabArena aims​​​‌ to evaluate models in‌ the settings that allow‌​‌ them to achieve peak​​ performance. This includes hyperparameter​​​‌ tuning with well-designed search‌ spaces, cross-validation, and ensembling‌​‌ different hyper-parameter configurations. Besides​​ providing an up-to-date comparison​​​‌ of models, TabArena provides‌ insights on the impact‌​‌ of cross-validation, ensembling, tuning,​​ and validation overfitting. Results,​​​‌ updated on a regular‌ basis with new methods,‌​‌ are presented on a​​ leaderboard at http://­tabarena.­ai.​​​‌

TabArena is reaching a‌ very broad visibility. Indeed,‌​‌ while it went public​​ only this summer, it​​​‌ is cited 26 times‌ 6 months later and‌​‌ received the spotlight distinction​​ at NeurIPS, the largest​​​‌ machine learning conference.

Learning with missing values‌​‌

A common practice for​​ handling missing values in​​​‌ tables consists in first‌ imputing missing values—i.e., replacing‌​‌ them with plausible values—and​​ then proceeding as if​​​‌ the data were complete.‌ In this context, we‌​‌ asked a simple but​​ fundamental question: is it​​​‌ worth investing effort and‌ resources in better imputations‌​‌ to improve predictions? This​​ work complements our previous​​​‌ asymptotic theoretical findings with‌ a thorough empirical finite-sample‌​‌ study 5, providing​​ useful conclusions for practitioners.​​​‌ Results show that better‌ recovery of missing values‌​‌ leads to better prediction,​​ but with diminishing returns:​​​‌ a large improvement in‌ recovery quality –which typically‌​‌ comes at a sizable​​ computational cost– leads to​​​‌ a small improvement in‌ prediction accuracy. The effect‌​‌ is further reduced when​​ using flexible learning algorithms,​​​‌ and adding missing-value indicators‌ Overall, on real-world datasets‌​‌ with powerful models, improving​​ imputation yields very limited​​​‌ benefits.

Guidance for evaluation‌ of medical AI

We‌​‌ contributed to a guidance​​ review on metric to​​​‌ evaluate predictors in the‌ context of medical practice‌​‌ 1. This guidance​​ is aimed at practitioners​​​‌ and is important given‌ the profusion of metrics‌​‌ applicable to classifiers, and​​ the confusions in what​​​‌ they measure. The work‌ outline both the various‌​‌ aspects that the metrics​​ probe –discrimination, calibration, overall​​​‌ performance, classification, and clinical‌ utility–, as well as‌​‌ the desirable mathematical properties.​​​‌ For instance, we stress​ that a good metric​‌ should be proper: it​​ should be optimal when​​​‌ the classifier outputs the​ true probability of events.​‌ The metrics are illustrated​​ in the context of​​​‌ medical usage, with an​ analysis of the utility​‌ and benefit to the​​ patient.

Double Debiased Machine​​​‌ Learning for Mediation Analysis​ with Continuous Treatments

We​‌ introduced double machine-learning estimators​​ with better convergence properties​​​‌ 43 to conduct a​ mediation analysis, ie​‌ to quantify how much​​ of the causal effect​​​‌ of a continuous treatment​ goes via an intermediate​‌ variable. We constructed a​​ kernel-based, Neyman-orthogonal estimator that​​​‌ combine regression and inverse-probability-weighting​ ideas while avoiding explicit​‌ estimation of the mediator​​ density, which is beneficial​​​‌ with high-dimensional or continuous​ mediators, that often occur​‌ in applications. We established​​ key theoretical properties: asymptotic​​​‌ normality at a nonparametric​ rate and multiple robustness​‌ that tolerates some misspecified​​ nuisance modelsand illustrate; derived​​​‌ an asymptotically mean-squared-error–optimal bandwidth​ and associated confidence intervals​‌ for the mediated response​​ curve. Simulation studies and​​​‌ an application to real-world​ medical data from the​‌ UK Biobak cohort (assessing​​ the mediating role of​​​‌ brain-related variables in the​ effect of glycemic control​‌ on cognitive outcomes) demonstrate​​ improved finite-sample performance over​​​‌ existing mediation estimators, highlighting​ the method’s practical relevance​‌ for complex observational studies.​​

Emergence of maths​​​‌ gender gap

Together with​ colleagues in cognitive psychology,​‌ we studied determinants of​​ the gender gap in​​​‌ mathematics abilities 6.​ We analyzed four consecutive​‌ cohorts of nation-wide evaluation​​ in France, on 5-to-7-year-old​​​‌ first graders. The data​ reveal the emergence of​‌ a gap in test​​ results during the first​​​‌ grade: girls and boys​ start the year with​‌ almost equal test performance,​​ but after one year​​​‌ of schooling the boys​ perform markedly better. This​‌ gender gap emerged across​​ all type of schooling​​​‌ (including Montessori or other​ innovative pedagogy), all family​‌ socio-economic status. The onset​​ of the gap was​​​‌ related to the admission​ in first grade, and​‌ not the age of​​ the children. In contrast​​​‌ to maths, the development​ of language skills follow​‌ different dynamics, with a​​ gap favoring girls present​​​‌ before schooling and different​ temporal evolution during schooling,​‌ narrowing this gap. The​​ study concludes that the​​​‌ gender gap is unlikely​ to be due to​‌ fundamental gender differences in​​ aptitudes, but rather likely​​​‌ mediated by interactions by​ teachers and parents, with​‌ hypothesis such as transmission​​ of anxiety or internalizing​​​‌ stereotypes.

Influence of training​ difficulty in learning outcomes​‌ of medical students

Literature​​ supports that in order​​​‌ to learn, tasks should​ not be too difficulty​‌ nor too easy. In​​ a study in press,​​​‌ we attempted to identify​ that optimal level of​‌ difficulty using millions of​​ student-question interactions of French​​​‌ students on the biggest​ medical training platform (​‌Banque nationale d'entraînement,​​ BNE) to determine how​​​‌ the difficulty of practice​ questions relative to student​‌ ability influences final exam​​ performance. The best learning​​ outcomes occur when students​​​‌ engage with questions that‌ are, on average, slightly‌​‌ easier than their current​​ proficiency level. This sweet​​​‌ spot for difficulty is‌ not universal; it varies‌​‌ significantly across different medical​​ specialties and individual student​​​‌ abilities. High-ability students, in‌ particular, showed greater sensitivity‌​‌ to question difficulty. These​​ results emphasize the need​​​‌ for adaptive learning systems‌ that can personalize difficulty‌​‌ in real-time to match​​ each student's evolving skills​​​‌ and the specific complexity‌ of the subject matter.‌​‌

Unpacking the scale narrative​​ in AI

Plotting the​​​‌ increase of the scale‌ of notable AI systems‌​‌ in the last years​​ reveals a staggering explosion.​​​‌ AI's size has been‌ growing super exponentially on‌​‌ a variety of dimensions:​​ training compute, training cost​​​‌ (fig:aiscale), inference cost, amount‌ of data used. Studying‌​‌ the wording used in​​ pivotal publications as well​​​‌ as company communications shows‌ that it anchors AI‌​‌ success in this growth,​​ thus settings implicit social​​​‌ norms around scale 8‌. But systematic analysis‌​‌ of benchmark results show​​ that scale does not​​​‌ always bring benefit. The‌ narrative of scale is‌​‌ simplified and leaves aside​​ many important ingredients of​​​‌ success of AI systems.‌ In addition, the race‌​‌ for scale comes with​​ planetary and societal consequences,​​​‌ which we study and‌ document 8. Ever-increasing‌​‌ inference costs threaten economic​​ and electricity sustainability. An​​​‌ unstoppable appetite for training‌ data leads to fitting‌​‌ models on enormous datasets​​ that elude quality control,​​​‌ engulfing undesirable facets of‌ internet (including child pornography)‌​‌ or eroding privacy. The​​ race for scale has​​​‌ financial consequences, benefiting above‌ all actors of compute,‌​‌ but also structuring an​​ ecosystem where cash-rich and​​​‌ GPU-rich actors have leverage‌ on priorities, industrial or‌​‌ academic. These actors sometimes​​ have circular investments strategies:​​​‌ funding third parties that‌ will spend all this‌​‌ funding in compute, which​​ can fuel an investment​​​‌ bubble in AI.

Evolution‌ of the training cost‌​‌ (in dollars) of notable​​ AI systems across the​​​‌ years

Evolution of the‌ training cost (in dollars)‌​‌ of notable AI systems​​ across the years

We conclude​​ our study, published at​​​‌ FAccT 8, by‌ underlining that academic research‌​‌ has a central role​​ to play in these​​​‌ dynamics and must shape‌ a healthy and grounded‌​‌ narrative. We recommend to​​ 1) pursue basic AI​​​‌ research of interest independent‌ of scale, eg uncertainty‌​‌ quantification, causality, etc. 2)​​ hold responsible norms, in​​​‌ particular avoiding asking for‌ compute increase when editing‌​‌ or reviewing, 3) always​​ publish measures of compute​​​‌ to document the tradeoffs.‌

This study has had‌​‌ much impact: it has​​ been well picked up​​​‌ by academics as well‌ as policy-makers, due to‌​‌ its relevance to the​​ current economy of innovation.​​​‌ It has been cited‌ 48 times in less‌​‌ than a year.

Going​​ from a theoretical causal​​​‌ analysis framework to practical‌ guidance with health data‌​‌

Many applications of machine​​ learning, in particular in​​​‌ healthcare, need to lead‌ to actionable conclusions and‌​‌ support for decision-making processes​​​‌ through. Thus, such applications​ must go beyond statistical​‌ associations and use a​​ causal framework that. This​​​‌ is challenging to implement​ in practice, particularly when​‌ dealing with noisy real-world​​ observational data. We propose​​​‌ and document a practical,​ five-step framework to turn​‌ routine electronic health records​​ (EHR) into reliable, causally-grounded​​​‌ evidence for treatment decisions​ 3, illustrated on​‌ the effect of albumin​​ plus crystalloids versus crystalloids​​​‌ alone on 28‑day mortality​ in sepsis. We emphasize​‌ that valid inference from​​ observational ICU data hinges​​​‌ on: (1) careful study​ design using a target-trial​‌ emulation/PICOT formulation to avoid​​ time-related biases such as​​​‌ immortal time bias; (2)​ explicit causal reasoning to​‌ identify confounders and define​​ an estimand; (3) robust​​​‌ estimation using modern causal​ estimators, where doubly robust​‌ methods with flexible machine-learning​​ nuisances (e.g. random forests)​​​‌ perform best; and (4)​ systematic “vibration” analyses to​‌ quantify how sensitive conclusions​​ are to design, confounder,​​​‌ and model choices. Applying​ this pipeline to MIMIC‑IV,​‌ they recover the “no​​ average effect” of albumin​​​‌ seen in randomized controlled​ trials (RCTs), while revealing​‌ clinically meaningful treatment heterogeneity,​​ with potential benefit in​​​‌ subgroups such as older​ patients, males, and those​‌ with septic shock, thereby​​ showcasing how valid causal​​​‌ machine learning on EHRs​ can complement RCTs for​‌ individualized decision-making.

Releases of scikit-learn​‌

2025 saw two major​​ releases of scikit-learn (1.7​​​‌ in June and 1.8​ in December). Scikit-learn has​‌ kept improving, adding both​​ user-visible features, and deep​​​‌ transformations of the technical​ piles. We list below​‌ a few highlights that​​ are certainly not exhaustive​​​‌ but illustrate the continuous​ progress made.

Here the​‌ user has expanded the​​ display for the LogisticRegression​​​‌ to reveal all the​ parameter values. Hovering on​‌ a parameter name reveals​​ the corresponding description.

           

HTML​​​‌ estimator display showing the​ parameter values

HTML estimator​‌ display showing the parameter​​ values

• Increasing​‌ support of GPUs
  We​​ are progressively rewriting the​​​‌ underlying compute operations to​ be able to execute​‌ on GPUs. As of​​ scikit-learn 1.8, full analyses​​​‌ can be run, including​ cross-validation and model evaluation.​‌ On many workflows, running​​ on the GPU leads​​​‌ to massive speedups (multiple​ folds, up to 70x).​‌
• Linear model speed ups​​
  The algorithmics of the​​​‌ linear models have been​ improved along many directions,​‌ leading to speed ups,​​ up to 10x in​​​‌ the sparse regression cases.​
• Temperature scaling recalibration
  Recalibration​‌ correct systematics biases in​​ prediction probabilities, eg over​​​‌ or under-confident classifiers. The​ problem becomes much harder​‌ in many class settings,​​ because each class comes​​​‌ with a probability that​ must be estimated. Temperature​‌ scaling is a recalibration​​ method that is particularly​​​‌ well suited to such​ settings.
• Estimator displays
  For​‌ a user working interactively​​ with scikit-learn, as most​​ data-scientists do, printing the​​​‌ models brings up a‌ rich display that we‌​‌ have been improving in​​ the last releases. The​​​‌ stakes are to make‌ the user more productive.‌​‌ As with all user-experience​​ work, the challenge is​​​‌ to display the right‌ information, and make it‌​‌ understandable. In the last​​ year, we have added​​​‌ a display of the‌ hyper-parameters of the estimator,‌​‌ as well as the​​ corresponding documentation, as illustrated​​​‌ in fig:estimatordisplay.
• Free threading‌
  The Python virtual machine‌​‌ has historically had a​​ central lock that prevented​​​‌ efficient thread-based parallel computing.‌ However, this lock has‌​‌ recently been removed and​​ the virtual machine can​​​‌ be built without it.‌ We have adapted scikit-learn‌​‌ to make sure that​​ it runs safely in​​​‌ heavily multi-threaded settings, opening‌ the door do data‌​‌ science in Python with​​ efficient native parallel computing.​​​‌

skrub

Skrub is a‌ package to facilitate machine‌​‌ learning on tables that​​ was first released at​​​‌ the end of 2023.‌ Year 2024 was a‌​‌ very active year for​​ skrub, with three release​​​‌ (0.5 in Jan, 0.6‌ in Jul, and 0.7‌​‌ in Dec), and the​​ following major features:

• DataOps​​​‌
  skrub now comes with‌ a new way of‌​‌ writing non-linear pipelines –dubbed​​ DataOps– that combine​​​‌ multiple tables, tracks provenance‌ through their transformations, and‌​‌ integrates machine learning. The​​ DataOps can then be​​​‌ re-applied to new data,‌ cross-validated, tuned, or extracted‌​‌ to be put in​​ production.
• Optuna
  In skrub​​​‌ 0.7, Optuna can be‌ used for hyper-parameter tuning‌​‌ on the pipelines. It​​ opens the door to​​​‌ advanced hyper-parameter optimization algorithms.‌

While skrub is a‌​‌ fairly new package, it​​ is increasingly well received​​​‌ by user. Uptake in‌ download numbers can be‌​‌ seen on pypistats.org/packages/skrub,​​ with, 6 000 downloads​​​‌ daily, as of end‌ of December, and a‌​‌ beautiful exponential growth.

joblib​​

joblib is a very​​​‌ simple computation engine in‌ Python that is massively‌​‌ used worldwide, including as​​ a dependency of packages​​​‌ such as scikit-learn for‌ parallel computing.

Release 1.5‌​‌. Many changes to​​ follow evolutions of the​​​‌ ecosystem and improve behaviors.‌ Major changes are:

• Avoiding‌​‌ collisions of cache when​​ cache is stored on​​​‌ a shared disk across‌ different nodes from a‌​‌ cluster
• Support of Python​​ 3.14

Probabl‌

Probabl is an Inria‌​‌ spin-off in which Gaël​​ Varoquaux has 30% of​​​‌ his time allocated and‌ is Chief Science Officer.‌​‌ Probabl's mission is to​​ develop and make sustainable​​​‌ an ecosystem of data-science‌ commons. Probabl is the‌​‌ larger employer of scikit-learn​​ maintainers. It builds a​​​‌ commercial offer around the‌ scikit-learn ecosystem by augmenting‌​‌ scikit-learn with solutions and​​ services for the entreprise.​​​‌ Gaël Varoquaux is the‌ point of contact at‌​‌ Soda.

Pass Culture

Within​​ the Ministry of Culture-Inria​​​‌ convention, Samuel Girard and‌ Jill-Jênn Vie have been‌​‌ involved in a partnership​​ with Pass Culture (used​​​‌ by 3 million students‌ in France) to improve‌​‌ the diversity of their​​​‌ recommendations (12 months, started​ in June 2024). We​‌ hired an engineer, Hiba​​ Bederina , from June​​​‌ 2024 from May 2025​ and conducted a randomized​‌ controlled trial on 400,000​​ users, which led to​​​‌ a publication on a​ RecSys workshop on social​‌ good.

Collaboration with Ministère​​ de la Santé

We​​​‌ have a 4-year long​ collaboration with Ministère de​‌ la Santé (HAS) on​​ using the national healthcare​​​‌ data for prevention and​ policy evaluation. Gaël Varoquaux​‌ and Judith Abecassis are​​ in charge at Soda.​​​‌

Collaboration with public​‌ interest group Pix

Jill-Jênn​​ Vie got a Paris​​​‌ Region PhD 2023 funding​ with Pix (certification of​‌ digital competencies, 6 million​​ active users), about optimizing​​​‌ human learning using reinforcement​ learning. Samuel Girard 's​‌ PhD is currently on​​ this funding (105,000 euros​​​‌ from région Île-de-France, 20,000​ euros from Pix).

10.1.1 Inria​​​‌ associate team not involved​ in an IIL or​‌ an international program

• Title:​​
  Recommendations Encouraging Diversity
• Duration:​​​‌
  2024 -> 2026
• Coordinator:​
  Jill Jenn Vie and​‌ Koh Takeuchi (takeuchi@i.kyoto-u.ac.jp)
• Partners:​​
  • Kyoto University (Japan)
  • CNRS​​​‌
• Inria contact:
  Jill Jenn​ Vie
• Summary:
  This project​‌ aims to create recommender​​ systems that optimize for​​​‌ cultural diversity. Finding items​ that not only optimize​‌ click-through rate, or profit,​​ but also encourage users​​​‌ to discover new things.​ The goal is to​‌ borrow methods from causal​​ inference to measure the​​​‌ treatment effect of recommendations​ (defined as the diversity​‌ after and before recommendation),​​ and methods from reinforcement​​​‌ learning to optimize this​ treatment effect. One key​‌ element to achieve this​​ project is that plenty​​​‌ of real data is​ available thanks to our​‌ current partnership with Pass​​ Culture, an app used​​​‌ by the French government​ to provide a budget​‌ ranging from 20 to​​ 300 euros for every​​​‌ 15 to 18 years​ old in order to​‌ purchase culture goods. These​​ works will be done​​​‌ between Soda team and​ Kyoto University, with the​‌ help of CNRS.

10.2.1​​​‌ Visits of international scientists​

10.2.2 Visits to international​‌ teams

10.3.1 Horizon Europe​

PEPR Santé Numérique​‌

Soda is part of​​ the “PEPR Santé Numérique”​​​‌ in the SMATCH subgroup​ that focuses on evidence​‌ of clinical efficacy. Soda​​ will address two questions.​​​‌ The first question, addressed​ in collaboration with the​‌ PreMedical team, is that​​ of external validity of​​​‌ randomized trials: how much​ is the treatment effect​‌ measured in a randomized​​ clinical trial affected by​​​‌ the sampling bias of​ the trial, the difference​‌ between the study population​​ and the intended target​​​‌ population. The second question,​ addressed in collaboration with​‌ the Heka team, is​​ that of defining guidelines​​​‌ to evaluate software as​ a medical device. One​‌ particular challenge that we​​ will tackle is to​​​‌ give procedures and recommendations​ to evaluate an update​‌ to a software used​​ in clinical decision making​​​‌ using historical data rather​ than a trial. The​‌ project started end of​​ 2023. Gaël Varoquaux is​​​‌ in charge at Soda,​ and Judith Abecassis is​‌ also supervising.

Project Partages​​

“Partages” is a large​​​‌ project funded by BPI​ France to develop digital​‌ commons for medical text​​ analysis. In particular, the​​​‌ project will create material​ suitable for fine-tuning or​‌ aligning language models to​​ perform best on French​​​‌ medical texts. Beyond the​ medical terms, there are​‌ specific challenges of clinical​​ texts: these often result​​​‌ from scanning notes that​ have been taken fast,​‌ full of context-specific abbreviations​​ and typos. The role​​​‌ of Soda is to​ design data-augmentation routine that​‌ help making language models​​ robust to these challenges.​​​‌ The project started end​ of 2024. Gaël Varoquaux​‌ is in charge at​​ Soda, and Judith Abecassis​​​‌ is also supervising.

ANR​ StatQA

Marine Le Morvan​‌ obtained an ANR JCJC​​ (2025-2029, 305 k€). LLMs​​​‌ provide unprecedented access to​ information, but their statistical​‌ reasoning abilities remain limited.​​ We introduce the concept​​​‌ of Statistical Question Answering​ (StatQA) to designate their​‌ capacity to address quantitative,​​ non-deterministic questions with calibrated​​ uncertainty. Our objectives are​​​‌ twofold: first, to assess‌ the statistical soundness of‌​‌ LLMs’ responses using institutional​​ datasets (INSEE, Eurostat); second,​​​‌ to develop multimodal approaches‌ that integrate tabular models‌​‌ with natural language. This​​ work aims to enhance​​​‌ the reliability and precision‌ of LLM outputs.

ANR‌​‌ TaFoMo

Gaël Varoquaux obtained​​ an ANR PRCE (2025-2029,​​​‌ 438 k€, partners Fabian‌ Suchanek at Telecom Paris‌​‌ and Antoine Neuraz at​​ Stane Group). The goal​​​‌ is to create Table‌ Foundation Models, pre-trained on‌​‌ large collections of tables,​​ embedding rich knowledge for​​​‌ subsequent machine-learning tasks. The‌ project involves 3 axis:‌​‌ 1) developing new architectures,​​ that handle different data​​​‌ types and multiple tables,‌ 2) pre-training models with‌​‌ diverse large data from​​ sources like Wikidata and​​​‌ DBpedia, drawing on PIs‌ expertise in databases and‌​‌ knowledge graphs, and 3)​​ and rigorously evaluating models​​​‌ across tasks, including health‌ applications, to confirm their‌​‌ practical value and robustness​​ to data variations.

ANR​​​‌ ICPC

Jill-Jênn Vie obtained‌ an ANR JCJC (2025-2029,‌​‌ 238 k€). The goal​​ is to develop an​​​‌ assistant to learn programming‌ by solving algorithmic problems‌​‌ like in coding contests.​​ We plan to generate​​​‌ hints without revealing the‌ solution, while exploring automatic‌​‌ testcase generation to break​​ an incorrect solution, to​​​‌ encourage robustness. We also‌ plan to generate or‌​‌ recommend exercises within the​​ proximal zone of development​​​‌ to keep students engaged.‌ The project will feature‌​‌ actual experiments of the​​ developed systems in classes,​​​‌ for example in high‌ school.

Conseil Scientifique CNIL​​

Gaël Varoquaux is a​​​‌ scientific expert at the‌ scientific committee of CNIL,‌​‌ the French data protection​​ authority.

11.1.1‌ Scientific events: organisation

11.1.2‌ Scientific events: selection

11.1.3 Journal

11.1.4 Invited​​​‌ talks

Gaël Varoquaux

• Académie‌ Royale de Médecine, Bélgique,‌​‌ journée sur l'IA et​​ la santé, Brussels
• Ellis-Helmoltz​​​‌ workshop on Foundation models‌ for science, Berlin
• End-to-end‌​‌ data processing workshop, sigmod,​​ Berlin
• Entente CordIAle Franco-English​​​‌ meetings, Cambridge
• Journée de‌ la santé de Santé,‌​‌ APHP, Paris
• Indaba Chad,​​ N'Djamena Chad (remote)
• Isaac​​​‌ Newton Institute, Cambridge
• Critéo‌ AI ethics days, Criteo,‌​‌ Paris
• Dagstuhl workshop on​​​‌ Table Representation Learning, Dagstuhl,​ Germany
• Dali workshop, Sorrento,​‌ Italy
• Python Exchange, remote​​
• ESMRMB keynote, Marseilles, France​​​‌
• EurIPS keynote, Copenhagen, Denmark​
• EurIPS benchmarking keynote, Copenhagen,​‌ Denmark
• Congrès de la​​ Société Française de Physique,​​​‌ Troyes, France
• NeurIPS in​ Paris keynote, Paris
• Séminaire​‌ Owkin
• Polish academy day​​ on AI in science,​​​‌ Paris
• P16 annual days,​ Paris
• PyData London Meeting,​‌ London
• Telecom Student association,​​ Saclay, France
• Teratec annual​​​‌ event keynote, Paris
• Journées​ de la SFDS sur​‌ l'incertitude, Paris
• VLDB panel​​ on tabular foundation models,​​​‌ London

David Holzmüller

• Group​ seminar, University of Amsterdam,​‌ Amsterdam, Netherlands
• Tabular foundation​​ models workshop, Freiburg, Germany​​​‌
• AutoML School 2025, Tübingen,​ Germany
• PriorLabs reading group,​‌ remote
• Group seminar, RWTH​​ Aachen, Aachen, Germany
• Group​​​‌ seminar, LMU München, Munich,​ Germany

Jill-Jênn Vie

• A​‌ Pre-Trained Graph-Based Model for​​ Adaptive Sequencing of Educational​​​‌ Documents, Kyoto University, Japan,​ January 29, 2025
• Efficiency​‌ and environmental impact of​​ LLMs, Inria Foresight Seminar,​​​‌ Rungis, March 26, 2025​

• A Pre-Trained Graph-Based Model​‌

  for Adaptive Sequencing of​​ Educational Documents, IRIT, Toulouse,​​​‌ July 7, 2025

• Optimal​ Training Difficulty for Optimizing​‌ Learning Outcomes, Saclay PhD​​ students day in STIC,​​​‌ Télécom Paris, Palaiseau, October​ 2, 2025

Judith Abécassis​‌

• VITE2025 : Explanability for​​ high-dimensional statistics, Montpellier
• Group​​​‌ Seminar, iPLesp, Paris
• Medical​ interns' seminar in Neurology​‌ at Lariboisière hospital, Paris​​
• Introduction to AI with​​​‌ EHR for anesthesia and​ intensive care residents, Paris​‌
• Paris Health AI Workshop,​​ Paris

Marine Le Morvan​​​‌

• Keynote at EurIPS'25 Workshop​ on AI for Tabular​‌ Data, Copenhagen, Denmark, December​​ 2025
• Keynote at Junior​​​‌ Conference on Data Sciences​ and Engineering, Paris, France,​‌ September 2025
• Probabilities and​​ statistics seminar, Laboratoire de​​​‌ Mathématiques d’Orsay, France, June​ 2025
• Table Representation Learning​‌ (TRL) seminar, ELLIS Unit​​ Amsterdam, Netherlands, April 2025​​​‌

11.1.5 Leadership within the​ scientific community

Gaël Varoquaux​‌

• Expert on the International​​ AI Safety Report 2025​​​‌

11.1.6 Scientific expertise

Gaël​ Varoquaux

• Reviewer for the​‌ general funding call at​​ ANR (AAPG)

Jill-Jênn Vie​​​‌

• Organisation internationale de la​ francophonie

Judith Abécassis

• Reviewer​‌ for the Messidore AAP​​ (Inserm)

Courses

• Gaël Varoquaux
   

  • Preparing​​ tabular data for machine​​​‌ learning, tutorial, EU ADS​ summer school, 3h, Luxembourg​‌
  • Health AI summer school,​​ Paris, France, 30 mn​​​‌
• Marine Le Morvan
   

  • APM_53441_EP​ - From Boosting to​‌ Foundation Models: learning with​​ Tabular Data, Ecole Polytechnique​​​‌ (Master 2), 30h
  • APM_51438_EP​ - Refresher Course in​‌ Artificial Intelligence, Ecole Polytechnique​​ (Master 1), 15h
  • Learning​​​‌ with missing values, Dauphine​ executive master, 6h
• Jill-Jênn​‌ Vie
   

  • Deep Learning, ENS​​ Paris, 27 h
  • CSC_41M02_EP​​​‌ Algorithms and Advanced Programming​ (ICPC training), École polytechnique,​‌ 18 h
  • SWERC training,​​ ENS Paris-Saclay, 30 h​​​‌ éq. TD
  • Tabular Deep​ Learning, Institut Polytechnique de​‌ Paris, 1 h
  • Computer​​ Vision, Ecole polytechnique, 45​​​‌ h
• Judith Abécassis
   

  • Causal​ Inference DS-UA 9201, NYU​‌ Paris, Spring 2025, 56h​​
  • AI for Healthcare, Centrale​​​‌ Supelec and Essec M2​ (Data Sciences & Business​‌ Analytics), 24h

11.2.1 Supervision​​

Gaël Varoquaux

• PhD supervision​​​‌
  • Jovan Stojanovic (50%), co-supervised​ with Margherita Comola (Paris​‌ School of Economics)
  • Julie​​ Alberge (30%), co-supervised with​​ Judith Abecassis (Soda, Inria)​​​‌
  • Sebastien Melo (30%), co-supervised‌ with Marine Le Morvan‌​‌ (Soda, Inria)
  • Celestin Eve​​ (50%), co-supervised with Thomas​​​‌ Moreau (Mind, Inria)
  • Meilame‌ Tayebjee (50%), co-supervised with‌​‌ Guillaume Lecué (ENSAE)
  • Félix​​ Lefevbre
  • Emma Cussenot, since​​​‌ December 2025 (25%), co-supervised‌ with Judith Abecassis (Soda,‌​‌ Inria) and Louis Potier​​ (AP-HP, Université Paris-Cité)
  • Gioia​​​‌ Blayer, since November 2025‌ (70%), co-supervised with Marine‌​‌ le Morvan (Soda, Inria)​​
• Internships
  • Emma Cussenot (50%,​​​‌ co-supervised with Judith Abecassis‌ (Soda, Inria)
  • Dan Suissa‌​‌ (30%, co-supervised with Judith​​ Abecassis (Soda, Inria)

Jill-Jênn​​​‌ Vie

• PhD supervision
  • Jean‌ Vassoyan (33%), co-supervised with‌​‌ Nicolas Vayatis
  • Samuel Girard​​ (33%), co-supervised with Amel​​​‌ Bouzeghoub
  • Marie Generali-Lince (33%),‌ co-supervised with Patrick Loiseau‌​‌ (FairPlay) and Solenne Gaucher​​ (École polytechnique)
• Internships
  • Anav​​​‌ Agrawal (L2), IIT Delhi‌

Judith Abécassis

• PhD supervision‌​‌
  • Julie Alberge (30%), co-supervised​​ with Gaël Varoquaux (Soda,​​​‌ Inria)
  • Emma Cussenot, since‌ December 2025 (25%), co-supervised‌​‌ with Louis Potier (AP-HP,​​ Université Paris-Cité) and Gaël​​​‌ Varoquaux (Soda, Inria)
  • Thaïs‌ Walter, since September 2025‌​‌ (50%), co-supervised with Jean-Damien​​ Ricard (AP-HP, Paris University)​​​‌
• Internships
  • Emma Cussenot (50%,‌ co-supervised with Gaël Varoquaux‌​‌ (Soda, Inria)
  • Dan Suissa​​ (70%, co-supervised with Gaël​​​‌ Varoquaux (Soda, Inria)
  • Guillaume‌ Bertho (33%), co-supervised with‌​‌ Adrien Coulet (HeKA, Inria)​​ and Eric Jouvent (AP-HP,​​​‌ Université Paris-Cité)

Marine Le‌ Morvan

• PhD supervision
  • Sebastien‌​‌ Melo (70%), co-supervised with​​ Gaël Varoquaux (Soda, Inria)​​​‌
  • Gioia Blayer (10%), co-supervised‌ with Gaël Varoquaux (Soda,‌​‌ Inria)
• Internships
  • Vlada Voronina​​ (70%), co-supervised with Oana​​​‌ Balalau (Cedar, INRIA)

11.2.2‌ Juries

Gaël Varoquaux

• PhD‌​‌ and HDR jury
  • PhD​​ Committee of Elena Albu,​​​‌ KU Leuven, Belgium
  • PhD‌ Committee of Arnaud Delaunoy,‌​‌ Université de Liège, Belgium​​
  • PhD Committee of Nicolas​​​‌ Hiebel, LISN Saclay, France‌
  • PhD Committee of Lawrence‌​‌ Steward, Inria Paris, France​​
  • PhD Committee of Charbel​​​‌ Kindji, Inria Rennes, France‌
  • HDR Committee of Cedric‌​‌ Gouypailler, CEA, France
• Jury​​ of the DataE grants​​​‌ from Ministère de la‌ Santé

Jill-Jênn Vie

• PhD‌​‌ midway committee
  • Loris Gaven,​​ Inria Bordeaux, France
  • Badmavasan​​​‌ Kirouchenassamy, Sorbonne University, France‌
  • Anass El-Ayady, Université de‌​‌ Lorraine, France
• Jury of​​ agrégation d'informatique
• Jury of​​​‌ École polytechnique entrance examinations‌

Judith Abécassis

• PhD midway‌​‌ committee
  • Wassila Khatir, Université​​ Côte d'Azur, France
  • Ala​​​‌ Eddine Boudemia, Sorbonne University,‌ France
• PhD jury
  • PhD‌​‌ Committee of Yannis Lombardi​​ (as examinatrice), Sorbonne University,​​​‌ France

Marine Le Morvan‌

• Jury for Associate Professor‌​‌ position in Statistics and​​ Machine Learning, Sorbonne Universit´e​​​‌ (Jussieu).

11.2.3 Educational and‌ pedagogical outreach

Gaël Varoquaux‌​‌

• Chroniqueur Les Échos: every​​ 3 months, a short​​​‌ article for the general‌ public around an AI‌​‌ topic
• Talk on AI​​ at the “amicale du​​​‌ corps des mines”
• Panel‌ on AI and health‌​‌ at the AI action​​ summit in Grand Palais​​​‌

Jill-Jênn Vie

• Risques et‌ opportunités de l'IA en‌​‌ éducation, formation des enseignants,​​ École supérieure d'ingénieurs Léonard​​​‌ de Vinci, Courbevoie, 10‌ avril 2025
• Research in‌​‌ personalized education, teaching competitive​​ programming, ENS Paris-Saclay, Gif-sur-Yvette,​​​‌ April 11, 2025
• Intelligence‌ artificielle, Algorithmique et‌​‌ programmation, CIRM (50 prep​​ school teachers in computer​​​‌ science), Marseille, May 8,‌ 2025
• Un système de‌​‌ recommandation de problèmes d'algo​​​‌ pour préparer Prologin et​ ICPC, Finals of Prologin​‌ Programming Contest 2025 (100​​ students under 20 years​​​‌ old), Le Kremlin-Bicêtre, May​ 31, 2025
• Systèmes de​‌ recommandation industriels et LLM​​ pour la recommandation, Online​​​‌ Pix Webinar, June 17,​ 2025
• Apprendre à l'heure​‌ de l'IA, Centre​​ Teilhard de Chardin, Orsay​​​‌ and online, November 27,​ 2025

11.3.1​‌ Participation in Live events​​

Gaël Varoquaux

• Talk on​​​‌ AI at an event​ for IT professionals at​‌ Lyon (Generation IA, ADIRA)​​
• Talk on AI and​​​‌ health at a general-public​ event organized at Antony​‌
• Talk on tabular AI​​ at the dotAI tech​​​‌ conference Antony
• Talk at​ BNP Paribas's data and​‌ AI annual event

Judith​​ Abécassis

• public recording of​​​‌ a public episode podcast​ "Nouvelles Héroïnes" at Inria​‌ Saclay, for the "Les​​ Rendez-vous des Jeunes Mathématiciennes​​​‌ et Informaticiennes (RJMI)" days​

11.3.2 Others science outreach​‌ relevant activities

Judith Abécassis​​

• Organization of Inria Women​​​‌ Lunches at Inria Saclay​

International‌ journals

• 10 articleJ.‌​‌Judith Abécassis, É.​​Élise Dumas, J.​​​‌Julie Alberge and G.‌Gaël Varoquaux. From‌​‌ prediction to prescription: Machine​​ learning and Causal Inference​​​‌.Annual Review of‌ Biomedical Data ScienceApril‌​‌ 2025HALDOI
• 11​​ articleR.Rishi Bommasani​​​‌, S.Sanjeev Arora‌, J.Jennifer Chayes‌​‌, Y.Yejin Choi​​, M.-F.Mariano-Florentino Cuéllar​​​‌, L.Li Fei-Fei‌, D. E.Daniel‌​‌ E Ho, D.​​Dan Jurafsky, S.​​​‌Sanmi Koyejo, H.‌Hima Lakkaraju, A.‌​‌Arvind Narayanan, A.​​Alondra Nelson, E.​​​‌Emma Pierson, J.‌Joelle Pineau, S.‌​‌Scott Singer, G.​​Gaël Varoquaux, S.​​​‌Suresh Venkatasubramanian, I.‌Ion Stoica, P.‌​‌Percy Liang and D.​​Dawn Song. Advancing​​​‌ science- and evidence-based AI‌ policy: Policy must be‌​‌ informed by, but also​​ facilitate the generation of,​​​‌ scientific evidence.Science‌3896759July 2025‌​‌, 459-461HALDOI​​back to text
• 12​​​‌ articleB.Ben van‌ Calster, G. S.‌​‌Gary S Collins,​​ A. J.Andrew J​​​‌ Vickers, L.Laure‌ Wynants, K. F.‌​‌Kathleen F Kerr,​​ L.Lasai Barreñada,​​​‌ G.Gael Varoquaux,‌ K.Karandeep Singh,‌​‌ K. G.Karel Gm​​ Moons, T.Tina​​​‌ Hernandez-Boussard, D.Dirk‌ Timmerman, D. J.‌​‌David J Mclernon,​​ M.Maarten van Smeden​​​‌ and E. W.Ewout‌ W Steyerberg. Evaluation‌​‌ of performance measures in​​ predictive artificial intelligence models​​​‌ to support medical decisions:‌ overview and guidance.‌​‌The Lancet Digital Health​​December 2025, 100916​​​‌HALDOI
• 13 article‌R.Riccardo Cappuzzo,‌​‌ A.Aimee Coelho,​​ F.Félix Lefebvre,​​​‌ P.Paolo Papotti and‌ G.Gaël Varoquaux.‌​‌ Retrieve, Merge, Predict: Augmenting​​​‌ Tables with Data Lakes​.Transactions on Machine​‌ Learning Research JournalMay​​ 2025HAL
• 14 article​​​‌S.Sylvie Delacroix,​ D.Diana Robinson,​‌ U.Umang Bhatt,​​ J.Jacopo Domenicucci,​​​‌ J.Jessica Montgomery,​ G.Gaël Varoquaux,​‌ C. H.Carl Henrik​​ Ek, V.Vincent​​​‌ Fortuin, Y.Yulan​ He, T.Tom​‌ Diethe, N.Neill​​ Campbell, M.Mennatallah​​​‌ El-Assady, S.Søren​ Hauberg, I.Ivana​‌ Dusparic and N. D.​​Neil D Lawrence.​​​‌ Beyond Quantification: Navigating Uncertainty​ in Professional AI Systems​‌.RSS: Data Science​​ and Artificial Intelligence1​​​‌1January 2025HAL​DOI
• 15 articleM.​‌Matthieu Doutreligne, T.​​Tristan Struja, J.​​​‌Judith Abecassis, C.​Claire Morgand, L.​‌ A.Leo Anthony Celi​​ and G.Gaël Varoquaux​​​‌. Step-by-step causal analysis​ of EHRs to ground​‌ decision-making.PLOS Digital​​ Health42February​​​‌ 2025, e0000721HAL​DOI
• 16 article M.​‌Matthieu Doutreligne and G.​​Gaël Varoquaux. How​​​‌ to select predictive models​ for decision making or​‌ causal inference? GigaScience 14​​ January 2025 HAL DOI​​​‌
• 17 articleR.Rosana​ El Jurdi, G.​‌Gaël Varoquaux and O.​​Olivier Colliot. Confidence​​​‌ intervals for performance estimates​ in brain MRI segmentation​‌.Medical Image Analysis​​103July 2025,​​​‌ 103565HALDOI
• 18​ articleS.-E.Stéphan-Eloïse Gras​‌ and G.Gaël Varoquaux​​. Connaître avec les​​​‌ modèles de langage :​ une rupture paradigmatique.​‌Intellectica - La revue​​ de l’Association pour la​​​‌ Recherche sur les sciences​ de la Cognition (ARCo)​‌81February 2025,​​ 85-110HAL
• 19 article​​​‌A.-S.Anne-Sophie Hamy,​ A.Agathe Chabassier,​‌ C.Clara Sebbag,​​ C.Christine Rousset-Jablonski,​​​‌ C.Clémentine Berkach,​ I.Isabelle Ray-Coquard,​‌ L.Laura Sablone,​​ L.Lauren Darrigues,​​​‌ E.Elise Dumas,​ A.Angélique Bobrie,​‌ W.William Jacot,​​ M.Marc Espié,​​​‌ S.Sylvie Giacchetti,​ F.Floriane Jochum,​‌ A.Aullène Toussaint,​​ G.Geneviève Plu-Bureau,​​​‌ L.Lorraine Maitrot-Mantelet,​ A.Anne Gompel,​‌ P.Paul Gougis,​​ R.Raphaëlle Bas,​​​‌ C.Christine Decanter,​ B.Bernard Asselain,​‌ C.Charles Coutant,​​ L.Lili Sohn,​​​‌ G.Guillemette Jacob,​ C.Claire Saule,​‌ S.Sophie Frank,​​ J.Judith Abécassis,​​​‌ F.Florence Coussy,​ F.Fabien Reyal and​‌ S.Seintinelles Research Network​​. Time-to-pregnancy in patients​​​‌ with previous breast cancer​ and unexposed women: a​‌ prospective exposed-unexposed cohort study​​.EClinicalMedicine86July​​​‌ 2025, 103392HAL​DOI
• 20 articleJ.-B.​‌Jean-Baptiste Julla, T.​​Théo Jolivet, C.​​​‌Candice Estellat, G.​Gaël Varoquaux, A.​‌Aurélie Carlier, J.-F.​​Jean-François Gautier, J.​​​‌Julie Alberge, Y.​Yawa Abouleka, A.​‌Audrey Bergès, E.​​Elise Liu, J.​​​‌Judith Abécassis, F.​Florence Tubach and L.​‌Louis Potier. Incidence​​ of death and amputation​​​‌ in patients with a​ first diabetic foot ulcer:​‌ results from the CODIA​​ cohort.Diabetes &​​ Metabolism516September​​​‌ 2025, 101700HAL‌DOI
• 21 articleM.‌​‌Marimuthu Kalimuthu, D.​​David Holzmüller and M.​​​‌Mathias Niepert. LOGLO-FNO:‌ Efficient Learning of Local‌​‌ and Global Features in​​ Fourier Neural Operators.​​​‌Transactions on Machine Learning‌ Research Journal2025HAL‌​‌DOI
• 22 articleE.​​ N.Erva Nihan Kandemir​​​‌, J.-J.Jill-Jênn Vie‌, A.Adam Sanchez-Ayte‌​‌, O.Olivier Palombi​​ and F.Franck Ramus​​​‌. Investigating the Influence‌ of Training Difficulty on‌​‌ the Learning Outcomes of​​ Medical Students.Journal​​​‌ of Computer Assisted Learning‌421February 2026‌​‌HALDOI
• 23 article​​M. J.Myung Jun​​​‌ Kim, F.Félix‌ Lefebvre, G.Gaëtan‌​‌ Brison, A.Alexandre​​ Perez-Lebel and G.Gaël​​​‌ Varoquaux. Table Foundation‌ Models: on knowledge pre-training‌​‌ for tabular learning.​​Transactions on Machine Learning​​​‌ Research JournalAugust 2025‌HAL
• 24 articleP.‌​‌Pauline Martinot, B.​​Bénédicte Colnet, T.​​​‌T. Breda, J.‌J. Sultan, L.‌​‌L. Touitou, P.​​P. Huguet, E.​​​‌Elizabeth Spelke, G.‌G. Dehaene-Lambertz, P.‌​‌P. Bressoux and S.​​Stanislas Dehaene. Rapid​​​‌ emergence of a maths‌ gender gap in first‌​‌ grade.Nature643​​8073June 2025,​​​‌ 1020 - 1029HAL‌DOI
• 25 articleC.‌​‌Clémence Réda, J.-J.​​Jill-Jênn Vie and O.​​​‌Olaf Wolkenhauer. Comprehensive‌ evaluation of pure and‌​‌ hybrid collaborative filtering in​​ drug repurposing.Scientific​​​‌ Reports152711January‌ 2025, 2711HAL‌​‌DOI
• 26 articleC.​​Clémence Réda, J.-J.​​​‌Jill-Jênn Vie and O.‌Olaf Wolkenhauer. Joint‌​‌ Embedding-Classifier Learning for Interpretable​​ Collaborative Filtering.BMC​​​‌ Bioinformatics261January‌ 2025, 26HAL‌​‌DOI
• 27 articleZ.​​Zacchary Sadeddine, W.​​​‌Winston Maxwell, G.‌Gaël Varoquaux and F.‌​‌ M.Fabian M. Suchanek​​. Large Language Models​​​‌ as Search Engines: Societal‌ Challenges.Sigir Forum‌​‌2025HAL
• 28 article​​T.Taehwan Yun,​​​‌ M. J.Myung Jun‌ Kim and H.Hyunjung‌​‌ Shin. Extremely fast​​ graph integration for semi-supervised​​​‌ learning via Gaussian fields‌ with Neumann approximation.‌​‌Pattern Recognition164August​​ 2025, 111495HAL​​​‌DOI

International peer-reviewed conferences‌

• 29 inproceedingsJ.Judith‌​‌ Abécassis, H.Houssam​​ Zenati, S.Sami​​​‌ Boumaïza, J.Julie‌ Josse and B.Bertrand‌​‌ Thirion. CO11.2 -​​ Explorer les fonctions cognitives​​​‌ dans UK Biobank avec‌ une analyse de médiation‌​‌ causale.EPICLIN 2025​​ - Conférence francophone d’EPIdémiologie​​​‌ CLINique73Bordeaux, France‌May 2025, 203025‌​‌HALDOI
• 30 inproceedings​​A.Anav Agrawal and​​​‌ J.-J.Jill-Jênn Vie.‌ AlgoAce: Retrieval-Augmented Generation for‌​‌ Assistance in Competitive Programming​​.Proceedings of 9th​​​‌ Educational Data Mining in‌ Computer Science Education Workshop‌​‌ (CSEDM 2025)CSEDM 2025​​ - 9th Educational Data​​​‌ Mining in Computer Science‌ Education WorkshopPalermo, Italy‌​‌July 2025HALDOI​​
• 31 inproceedingsJ.Julie​​​‌ Alberge, V.Vincent‌ Maladière, O.Olivier‌​‌ Grisel, J.Judith​​ Abécassis and G.Gaël​​​‌ Varoquaux. Survival Models:‌ Proper Scoring Rule and‌​‌ Stochastic Optimization with Competing​​​‌ Risks.Proceedings of​ the 28th International Conference​‌ on Artificial Intelligence and​​ Statistics (AISTATS) 2025, Mai​​​‌ Khao, Thailand. PMLR: Volume​ 258.AISTATS 2025 -​‌ 28th International Conference on​​ Artificial Intelligence and Statistics​​​‌Phuket, ThailandMay 2025​HALback to text​‌
• 32 inproceedingsP.Pascaline​​ André, C.Charles​​​‌ Heitz, E.Evangelia​ Christodoulou, A.Annika​‌ Reinke, C. H.​​Carole H Sudre,​​​‌ M.Michela Antonelli,​ M. J.M Jorge​‌ Cardoso, A.Antoine​​ Gilson, S.Sophie​​​‌ Tezenas Du Montcel,​ G.Gaël Varoquaux,​‌ L.Lena Maier-Hein and​​ O.Olivier Colliot.​​​‌ Some hidden traps of​ confidence intervals in medical​‌ image segmentation: coverage issues​​.Lecture Notes in​​​‌ Computer ScienceBRIDGE 2025​ - MICCAI Workshop Bridging​‌ Regulatory Science and Medical​​ Imaging EvaluationMICCAI Workshops​​​‌Deajeon, South KoreaSeptember​ 2025HAL
• 33 inproceedings​‌D.Daniel Beaglehole,​​ D.David Holzmüller,​​​‌ A.Adityanarayanan Radhakrishnan and​ M.Mikhail Belkin.​‌ xRFM: Accurate, scalable, and​​ interpretable feature learning models​​​‌ for tabular data.​AITD 2025 – Workshop​‌ on AI for Tabular​​ DataCopenhagen, Denmark2025​​​‌HALDOI
• 34 inproceedings​S.Samuel Girard,​‌ J. D.Juan D​​ Pinto, J.-J.Jill-Jênn​​​‌ Vie and A.Amel​ Bouzeghoub. RegKT: Interpretable​‌ and Robust Deep Knowledge​​ Tracing With IRT-Regularizer.​​​‌2nd Human-Centric eXplainable AI​ in Education (HEXED) Workshop​‌Palermo, ItalyJuly 2025​​HAL
• 35 inproceedingsC.​​​‌Carole Ibrahim, H.​Hiba Bederina, C.​‌Cuesta Daniel, M.​​Montier Laurent, D.​​​‌Delabre Cyrille and J.-J.​Jill-Jênn Vie. Diversified​‌ recommendations of cultural activities​​ with personalized determinantal point​​​‌ processes.RecSoGood 2025​ - Second International Workshop​‌ on Recommender Systems for​​ Sustainability and Social Good​​​‌Prague, Czech RepublicSeptember​ 2025HAL
• 36 inproceedings​‌M.Marine Le Morvan​​ and G.Gaël Varoquaux​​​‌. Imputation for prediction:​ beware of diminishing returns​‌.International Conference on​​ Learning RepresentationsICLR 2025​​​‌ - International Conference on​ Learning RepresentationsSingapore, Singapore​‌April 2025HAL
• 37​​ inproceedingsF.Félix Lefebvre​​​‌ and G.Gaël Varoquaux​. Scalable Feature Learning​‌ on Huge Knowledge Graphs​​ for Downstream Machine Learning​​​‌.Neural Information Processing​ SystemsNeurIPS 2025 -​‌ 39th Annual Conference on​​ Neural Information Processing Systems​​​‌San Diego (California), United​ StatesDecember 2025HAL​‌
• 38 inproceedingsA.Alexandre​​ Perez-Lebel, G.Gaël​​​‌ Varoquaux, S.Sanmi​ Koyejo, M.Matthieu​‌ Doutreligne and M.Marine​​ Le Morvan. Decision​​​‌ from Suboptimal Classifiers: Excess​ Risk Pre-and Post-Calibration.​‌Proceedings of the 28th​​ International Conference on Artificial​​​‌ Intelligence and Statistics (AISTATS)​ 2025, Mai Khao, Thailand.​‌ PMLR: Volume 258.AISTATS​​ 2025 - the 28th​​​‌ International Conference on Artifi-​ cial Intelligence and Statistics​‌Mai Khao, ThailandMay​​ 2025HAL
• 39 inproceedings​​​‌R.Roman Plaud,​ A.Alexandre Perez-Lebel,​‌ M.Matthieu Labeau,​​ A.Antoine Saillenfest and​​​‌ T.Thomas Bonald.​ To Each Metric Its​‌ Decoding: Post-Hoc Optimal Decision​​ Rules of Probabilistic Hierarchical​​​‌ Classifiers.ICML 2025​ - 42nd International Conference​‌ on Machine LearningVancouver​​ (CA), CanadaJuly 2025​​HAL
• 40 inproceedingsJ.​​​‌Jingang Qu, D.‌David Holzmüller, G.‌​‌Gaël Varoquaux and M.​​Marine Le Morvan.​​​‌ TabICL: A Tabular Foundation‌ Model for In-Context Learning‌​‌ on Large Data.​​ICML 2025 - 42nd​​​‌ International Conference on Machine‌ LearningVancouver, CanadaJuly‌​‌ 2025HAL
• 41 inproceedings​​M.Moreno Ursino,​​​‌ S.Sandrine Boulet,‌ C.Corinne Collignon,‌​‌ F.Florence Francis-Oliviero,​​ E.Edouard Lhomme,​​​‌ R.Raphaël Porcher,‌ F.Florence Saillour,‌​‌ G.Gaël Varoquaux,​​ V.Vincent Vercamer,​​​‌ R.Rodolphe Thiébaut and‌ S.Sarah Zohar.‌​‌ Innovative clinical trial approach​​ for evaluating digital medical​​​‌ devices under European HTA‌ fast-Track frameworks.ISCB‌​‌ 2025 - 46th Annual​​ Conference of the International​​​‌ Society for Clinical Biostatistics‌Basel, SwitzerlandAugust 2025‌​‌HAL
• 42 inproceedingsG.​​Gaël Varoquaux, A.​​​‌ S.Alexandra Sasha Luccioni‌ and M.Meredith Whittaker‌​‌. Hype, Sustainability, and​​ the Price of the​​​‌ Bigger-is-Better Paradigm in AI‌.FAccT 2025 -‌​‌ ACM Conference on Fairness,​​ Accountability, and TransparencyAthens,​​​‌ GreeceJuly 2025HAL‌back to text
• 43‌​‌ inproceedingsH.Houssam Zenati​​, J.Judith Abécassis​​​‌, J.Julie Josse‌ and B.Bertrand Thirion‌​‌. Double Debiased Machine​​ Learning for Mediation Analysis​​​‌ with Continuous Treatments.‌Proceedings of Machine Learning‌​‌ ResearchAISTATS - 28th​​ International Conference on Artificial​​​‌ Intelligence and StatisticsPMLR-‌Mai Khao, ThailandMay‌​‌ 2025HALback to​​ text

Conferences without proceedings​​​‌

• 44 inproceedingsS.Sandrine‌ Boulet, V.Vincent‌​‌ Thévenet, J.Judith​​ Abécassis, C.Corinne​​​‌ Collignon, B.Bruno‌ Giraudeau, E.Edouard‌​‌ Lhomme, F.Francois​​ Petit, R.Raphaël​​​‌ Porcher, L.Laura‌ Richert, F.Florence‌​‌ Saillour, M.Moreno​​ Ursino, G.Gaël​​​‌ Varoquaux, V.Vincent‌ Vercamer, S.Sarah‌​‌ Zohar and R.Rodolphe​​ Thiébaut. Narrative review​​​‌ on the clinical evaluation‌ of AI-based digital medical‌​‌ devices from a methodological​​ perspective.PEPR-SanteNum 2025​​​‌ - Journées Annuelles du‌ PEPR Santé NumériqueLille,‌​‌ FranceOctober 2025HAL​​
• 45 inproceedingsN.Nick​​​‌ Erickson, L.Lennart‌ Purucker, A.Andrej‌​‌ Tschalzev, D.David​​ Holzmüller, P. M.​​​‌Prateek Mutalik Desai,‌ D.David Salinas and‌​‌ F.Frank Hutter.​​ TabArena: A Living Benchmark​​​‌ for Machine Learning on‌ Tabular Data.NeurIPS‌​‌ 2025 - The Thirty-ninth​​ Annual Conference on Neural​​​‌ Information Processing Systems Datasets‌ and Benchmarks TrackSan‌​‌ Diego, United States2025​​HAL
• 46 inproceedingsF.​​​‌Félix Lefebvre, M.‌ J.Myung Jun Kim‌​‌ and G.Gaël Varoquaux​​. Knowledge-Rich Embeddings for​​​‌ Tabular Learning.AITD‌ 2025 - Workshop on‌​‌ AI for Tabular Data​​Copenhagen, DenmarkDecember 2025​​​‌HAL
• 47 inproceedingsD.‌Daniel Musekamp, M.‌​‌Marimuthu Kalimuthu, D.​​David Holzmüller, M.​​​‌Makoto Takamoto and M.‌Mathias Niepert. Active‌​‌ Learning for Neural PDE​​ Solvers.ICLR 2025​​​‌ - The Thirteenth International‌ Conference on Learning Representations‌​‌Singapore, Singapore2025HAL​​DOI

Reports & preprints​​​‌

• 48 miscJ.Judith‌ Abécassis, H.Houssam‌​‌ Zenati, S.Sami​​​‌ Boumaïza, J.Julie​ Josse and B.Bertrand​‌ Thirion. Causal mediation​​ analysis with one or​​​‌ multiple mediators: a comparative​ study.May 2025​‌HAL
• 49 reportY.​​Yoshua Bengio, S.​​​‌Sören Mindermann, D.​Daniel Privitera, T.​‌Tamay Besiroglu, R.​​Rishi Bommasani, S.​​​‌Stephen Casper, Y.​Yejin Choi, P.​‌Philip Fox, B.​​Ben Garfinkel, D.​​​‌Danielle Goldfarb, H.​Hoda Heidari, A.​‌Anson Ho, S.​​Sayash Kapoor, L.​​​‌Leila Khalatbari, S.​Shayne Longpre, S.​‌Sam Manning, V.​​Vasilios Mavroudis, M.​​​‌Mantas Mazeika, J.​Julian Michael, J.​‌Jessica Newman, K.​​ Y.Kwan Yee Ng​​​‌, C.Chinasa Okolo​, D.Deborah Raji​‌, G.Girish Sastry​​, E.Elizabeth Seger​​​‌, T.Theodora Skeadas​, T.Tobin South​‌, E.Emma Strubell​​, F.Florian Tramèr​​​‌, L.Lucia Velasco​, N.Nicole Wheeler​‌, D.Daron Acemoglu​​, O.Olubayo Adekanmbi​​​‌, D.David Dalrymple​, T.Thomas Dietterich​‌, E.Edward Felten​​, P.Pascale Fung​​​‌, P.-O.Pierre-Olivier Gourinchas​, F.Fredrik Heintz​‌, G.Geoffrey Hinton​​, N.Nick Jennings​​​‌, A.Andreas Krause​, S.Susan Leavy​‌, P.Percy Liang​​, T.Teresa Ludermir​​​‌, V.Vidushi Marda​, H.Helen Margetts​‌, J.John Mcdermid​​, J.Jane Munga​​​‌, A.Arvind Narayanan​, A.Alondra Nelson​‌, C.Clara Neppel​​, A.Alice Oh​​​‌, G.Gopal Ramchurn​, S.Stuart Russell​‌, M.Marietje Schaake​​, B.Bernhard Schölkopf​​​‌, D.Dawn Song​, A.Alvaro Soto​‌, L.Lee Tiedrich​​, G.Gaël Varoquaux​​​‌, A.Andrew Yao​, Y.-Q.Ya-Qin Zhang​‌, F.Fahad Albalawi​​, M.Marwan Alserkal​​​‌, O.Olubunmi Ajala​, G.Guillaume Avrin​‌, C.Christian Busch​​, A. C.André​​​‌ Carlos Ponce de Leon​ Ferreira de Carvalho,​‌ B.Bronwyn Fox,​​ A. S.Amandeep Singh​​​‌ Gill, A. H.​Ahmet Halit Hatip,​‌ J.Juha Heikkilä,​​ G.Gill Jolly,​​​‌ Z.Ziv Katzir,​ H.Hiroaki Kitano,​‌ A.Antonio Krüger,​​ C.Chris Johnson,​​​‌ S.Saif Khan,​ K. M.Kyoung Mu​‌ Lee, D. V.​​Dominic Vincent Ligot,​​​‌ O.Oleksii Molchanovskyi,​ A.Andrea Monti,​‌ N.Nusu Mwamanzi,​​ M.Mona Nemer,​​​‌ N.Nuria Oliver,​ J. R.José Ramón​‌ López Portillo, B.​​Balaraman Ravindran, R.​​​‌ P.Raquel Pezoa Rivera​, H.Hammam Riza​‌, C.Crystal Rugege​​, C.Ciarán Seoighe​​​‌, J.Jerry Sheehan​, H.Haroon Sheikh​‌, D.Denise Wong​​ and Y.Yi Zeng​​​‌. International AI Safety​ Report.AI safety​‌ institute2025HALback​​ to text
• 50 misc​​​‌C.Clément Berenfeld,​ A.Ahmed Boughdiri,​‌ B.Bénédicte Colnet,​​ W. A.Wouter A.​​​‌ C. van Amsterdam,​ A.Aurélien Bellet,​‌ R.Rémi Khellaf,​​ E.Erwan Scornet and​​ J.Julie Josse.​​​‌ Causal Meta-Analysis: Rethinking the‌ Foundations of Evidence-Based Medicine‌​‌.May 2025HAL​​
• 51 miscE.Eugene​​​‌ Berta, D.David‌ Holzmüller, M. I.‌​‌Michael I. Jordan and​​ F.Francis Bach.​​​‌ Rethinking Early Stopping: Refine,‌ Then Calibrate.January‌​‌ 2025HAL
• 52 misc​​E.Eugène Berta,​​​‌ D.David Holzmüller,‌ M. I.Michael I‌​‌ Jordan and F.Francis​​ Bach. Structured Matrix​​​‌ Scaling for Multi-Class Calibration‌.November 2025HAL‌​‌
• 53 miscR.Rahul​​ Bordoloi, C.Clémence​​​‌ Réda, S.Saptarshi‌ Bej and O.Olaf‌​‌ Wolkenhauer. Handling Missing​​ Data in Downstream Tasks​​​‌ With Distribution-Preserving Guarantees.‌2025HAL
• 54 misc‌​‌S.Sacha Braun,​​ D.David Holzmüller,​​​‌ M. I.Michael I.‌ Jordan and F.Francis‌​‌ Bach. Conditional Coverage​​ Diagnostics for Conformal Prediction​​​‌.December 2025HAL‌
• 55 miscF.Fateme‌​‌ Ghayem, R.Raphael​​ Meudec, J.Jérôme​​​‌ Dockès, B.Bertrand‌ Thirion and D.Demian‌​‌ Wassermann. NeuroConText: Contrastive​​ learning for neuroscience meta-analysis​​​‌ with rich text representation‌.May 2025HAL‌​‌DOI
• 56 miscD.​​David Holzmüller and M.​​​‌Max Schölpple. Beyond‌ ReLU: How Activations Affect‌​‌ Neural Kernels and Random​​ Wide Networks.June​​​‌ 2025HALDOI
• 57‌ miscS.Sébastien Melo‌​‌, G.Gaël Varoquaux​​ and M.Marine Le​​​‌ Morvan. Epistemic Uncertainty‌ Quantification to Improve Decisions‌​‌ From Black-Box Models.​​December 2025HALback​​​‌ to text
• 58 misc‌R.Raphael Meudec,‌​‌ J.Jérôme Dockès,​​ F.Fateme Ghayem,​​​‌ D.Demian Wassermann and‌ B.Bertrand Thirion.‌​‌ Peaks2Image: reconstructing fMRI maps​​ from stereotactic coordinates to​​​‌ enhance cognitive meta-analysis.‌August 2025HAL

Other‌​‌ scientific publications

• 59 inproceedings​​J.Julie Alberge,​​​‌ V.Vincent Maladière,‌ O.Olivier Grisel,‌​‌ J.Judith Abécassis and​​ G.Gaël Varoquaux.​​​‌ Survival models: Proper scoring‌ rule and stochastic optimization‌​‌ with competing risks.​​EPICLIN 2025 - conférence​​​‌ francophone d’ÉPIdémiologie CLINique73‌Bordeaux, FranceMay 2025‌​‌, 203083HALDOI​​
• 60 inproceedingsS.Sandrine​​​‌ Boulet, V.Vincent‌ Thévenet, J.Judith‌​‌ Abécassis, C.Corinne​​ Collignon, B.Bruno​​​‌ Giraudeau, E.Edouard‌ Lhomme, F.Francois‌​‌ Petit, R.Raphaël​​ Porcher, L.Laura​​​‌ Richert, F.Florence‌ Saillour, M.Moreno‌​‌ Ursino, G.Gaël​​ Varoquaux, V.Vincent​​​‌ Vercamer, S.Sarah‌ Zohar and R.Rodolphe‌​‌ Thiébaut. Narrative review​​ on the clinical evaluation​​​‌ of AI-based digital medical‌ devices from a methodological‌​‌ perspective.SbN 2025​​ - 10th Statistics &​​​‌ Biopharmacy ConferenceParis, France‌October 2025HAL
• 61‌​‌ inproceedingsS.Sandrine Boulet​​, V.Vincent Thévenet​​​‌, J.Judith Abécassis‌, C.Corinne Collignon‌​‌, B.Bruno Giraudeau​​, E.Edouard Lhomme​​​‌, F.Francois Petit‌, R.Raphaël Porcher‌​‌, L.Laura Richert​​, F.Florence Saillour​​​‌, M.Moreno Ursino‌, G.Gaël Varoquaux‌​‌, V.Vincent Vercamer​​, S.Sarah Zohar​​​‌ and R.Rodolphe Thiébaut‌. Narrative review on‌​‌ the clinical evaluation of​​​‌ AI-based digital medical devices​ from a methodological perspective​‌.PEPR-SanteNum 2025 -​​ Journées Annuelles du PEPR​​​‌ Santé NumériqueLille, France​October 2025HAL
• 62​‌ inproceedingsM.Moreno Ursino​​, S.Sandrine Boulet​​​‌, C.Corinne Collignon​, F.Florence Francis-Oliviero​‌, E.Edouard Lhomme​​, R.Raphaël Porcher​​​‌, F.Florence Saillour​, G.Gaël Varoquaux​‌, R.Rodolphe Thiébaut​​ and S.Sarah Zohar​​​‌. Innovative Clinical Trial​ Approach for Evaluating Digital​‌ Medical Devices under European​​ Fast-Track Regulatory Frameworks.​​​‌EPICLIN 2025 - Conférence​ francophone d’EPIdémiologie CLINiqueBordeaux,​‌ FranceMay 2025HAL​​