Les lois contraignant les véhicules lourds pour protéger les véhicules légers (stationnement, distance de déplacement) ne sont pas appliquées, des forces de l'ordre refusant de verbaliser. Mais on invente de nouvelles lois contraignant les véhicules légers. ⚠ aux verbalisations à sens unique

To know more about the position and the corresponding research please read the full description, and follow the instructions to apply team.inria.fr/soda/files/2...

You're only a real statisticien if you can say "Heteroscedasticity" without flinching

Modern AIs tackle different questions than data science on industry or scientific applications. Statistical thinking, front and center in data science, is often hidden in AI. But it’s just as crucial. And too often, we treat data science or AI as merely a programming exercise

We can apply this to choosing which LLM to use to get correct answers, but with small costs: Start with small model Evaluate its epistemic uncertainty for a query If epistemic is high: a better answer can be given, call a more powerful model 5/

The suboptimality depends on the costs of false detections and misses. These costs define an optimal threshold. The more the classifier probabilities explore the region where predicted probability falls on the wrong side of the optimal threshold, the higher the suboptimality. 4/

Some maths 😀 gives us a point-wise (ie individual) estimate on how suboptimal a classifier- based decision is. 3/

The intuition: we find regions that drive heterogeneity in a classifiers error, and we characterize calibration on these. Such variance in probabilistic errors is actually "grouping" loss, and helps fully characterizing epistemic error, ie how suboptimal a classifier is 2/

#ICLR2026 paper✨️: Quantifying epistemic uncertainty of Blackbox classifiers, and link to better decisions Calibration on steroids, qualifying full prediction uncertainty with no need for Bayes, and tuning individual decisions 👇

Skore's tagline is “track your data science”. We’re relentlessly improving the tools, and we will push them to redefine the data-science experience. We’re opening early, because at Probabl we believe in transparency. This is just the beginning🔥

Skore complements building blocks such as @scikit-learn.org, helping to track and validate the work. It comes both as a powerful stand-alone open-source library, which can be used fully offline, and an online platform to track and share months of iterations docs.skore.probabl.ai/stable/auto_...

TabICLv2 is the next big step in foundation models for tabular data. Try the package, read the paper. Enjoy! Installation: pypi.org/project/tabi... Preprint: arxiv.org/abs/2602.11139 Open source code: github.com/soda-inria/t...

Looking beyond the median cost per sample, compute costs grow non-linearly with the data size. Figure 6 shows the influence of sample size, revealing the marked benefit of TabICLv2 over TabPFN-2.5, that becomes more and more pronounced with sample size. And TabICLv2 runs well on CPU.

TabICLv2 is state of the art, and fast. Figure 1 in the paper gives the results on the standard TabArena benchmark, showing that TabICLv2 is the best predictor without needing any hyper-parameter tuning.

🎉 Announcing TabICLv2: State-of-the art Table Foundation Model, fast and open source A breakthrough for tabular ML: better prediction and faster runtime than alternatives, work by Jingang Qu, David Holzmüller @dholzmueller.bsky.social , Marine Le Morvan, and myself 👇

Debuging a @pola.rs memory leak, metallica version *** But the memory remains ... Hashes to hashes Rust to rust Fade to black (no, to ruff!)

Finally did it! A glitter @scikit-learn.org sticker on my bike helmet. If you're around Saclay, Massy, ... and you see me go by on my bike, go ahead and cheer!

Morning ski A picture of an early-morning ski-touring run. I liked the shape drawn in the sky by the clouds flickr.com/photos/gaelv... Click through for high-res 👆

If someone with communication and/or influence training wants to help me out, I'm writing a piece on a conflict on which I would like not to make image-related mistakes. Do reach out. Thanks!

PPS: #EurIPS was amazing, great vibes and high quality science.

Returning from #EurIPS via a long-distance night bus rather than airplane for carbon footprint reason. I think it is important that even senior researchers change our behavior: practices of the elite pull consumption up, asxour reaction to opportunities set norms. PS: I do find it painful 😄

Super excited (and a bit nervous) to give a keynote tomorrow afternoon at #EurIPS. I'll be math, LLMs, and politics 🔥

Found a sharpie, thanks Christine Funke from @ellis.eu

#EurIPS Poster 81: what huge graph embedding for feature learning taught us Failed to print the poster, but happy to talk. If you bring me paper and sharpies, will draw 😄 eurips.cc/conference/p...

We also learned that performance on link prediction, the canonical task of knowledge-graph embedding, is not a good proxy for downstream utility. We believe this is because link prediction only needs local structure, unlike downstream tasks 9/10

Our approach, SEPAL, combines these elements for feature learning on large knowledge graphs. It creates feature vectors that lead to better performance on downstream tasks, and it is more scalable. Larger knowledge graphs give feature vectors that provide downstream value 8/10

Splitting huge knowledge graphs in sub-parts is actually hard because of the mix of very highly-connected nodes, and a huge long tail hard to reach. We introduce a procedure that allows for overlap in the blocks, relaxing a lot the difficulty. 7/10

To have a very efficient algorithm, we split the graph in overlapping highly-connected blocks that fit in GPU memory. Propagation is then simple in-memory iterations, and we embed huge graphs on a single GPU. 6/10

Knowledge graphs have long-tailed entity distributions, with many weakly-connected entities on which contrastive learning is under constrained. For these, we propagate embeddings via the relation operators, in a diffusion-like step, extrapolating from the central entities. 5/10