Éloi Zablocki

Senior research scientist at valeo.ai

photo_eloi_zablocki.jpg

I am a senior research scientist at valeo.ai, working on:

  • autonomous driving, scene understanding and forecasting, world-models, motion planning
  • vision and language, explainability, foundation models

Before joining Valeo, I completed a Ph.D. at Sorbonne Université in 2019 on multi-modal machine learning with language and vision, supervised by Patrick Gallinari, Laure Soulier, and Benjamin Piwowarski.

Prior to this, I earned an Engineering Degree from École Polytechnique (X2012) and an MSc from ENS Paris-Saclay (“Master MVA”).

News

Aug 2026 How Far Can 5,500 Hours of Driving Take You? is accepted at the ECCV DriveX Workshop 2026 (Oral). Guided by scaling laws established from 1M to 1B parameters, we release a 9B-parameter controllable world model for driving, open-weights and with the full recipe to build one, trained from scratch on 5,500 hours of driving. It sets the state of the art for driving video generation.
Aug 2026 New preprint: Pictura is a GPU-accelerated multi-agent driving simulator that renders every agent’s egocentric view at each step, sustaining up to 500K agent-steps/s on a single H100. With it, we train Alberti by self-play with plain PPO over 50B agent steps (~35M km): the first large-scale driving self-play policy learned directly from perspective images, with no privileged observation of the surroundings.
Jul 2026 New preprint: RDM trains a one-step image generator by matching the distribution of frozen encoder embeddings between generated and real images, with no teacher, discriminator, or denoising trajectory. It sets the one-step ImageNet state of the art and turns four-step FLUX.2 [klein] into a single step.
Jun 2026 New preprint: TOAD optimizes driving trajectories at test time by treating the planner’s scorer as a learned reward, improving end-to-end planners.
Apr 2026 DRIV-EX is accepted at ACL (Findings) 2026. Biasing autoregressive decoding with gradient-optimized embeddings enables the generation of counterfactual explanations for the LLMs used in autonomous driving.
Mar 2026 I am serving as an Area Chair for ECCV 2026.
Feb 2026 NAF is accepted at CVPR 2026 (Highlight). Image-guided neighborhood attention provides a lightweight and fast feature upsampler that generalizes zero-shot across vision foundation models while surpassing specialized architectures.
Feb 2026 MAD is accepted at CVPR 2026. Decoupling motion learning from appearance synthesis enables efficient adaptation of general video diffusion models into controllable, state-of-the-art driving world models with minimal supervision.
Feb 2026 DrivoR is accepted at CVPR 2026. Compressing multi-camera ViT features into a few register tokens enables a simple pure-transformer planner to achieve state-of-the-art end-to-end driving in both open and closed-loop settings.
Feb 2026 Loick Chambon defends his PhD entitled “Efficient Representations for Autonomous Driving”. Jury: Alexandre Alahi, Vincent Lepetit, Fatma Güney, Catherine Achard, Matthieu Cord, and myself. Congrats Loick!
Feb 2026 GIFT :gift: is accepted at TMLR 2026, with a Featured Certification. A framework for generating global, interpretable textual explanations of vision classifiers, combining counterfactual visual explanations with VLMs and LLMs.
Jan 2026 PPT is accepted at ICRA 2026. Pseudo-labeled trajectories can be used to pre-train trajectory prediction models: improved performance, efficiency, and generalization.

Current Students

Alumni

Scientific Service

Area Chair: ECCV 2026

Reviewer:
Conferences: CVPR 2021–2026, ECCV/ICCV 2021–2025, ACL 2026, ICLR 2025, ICML 2025, WACV 2024, IROS 2022, AAAI 2021
Journals: IJCV 2022,2025, TPAMI 2021–2022, T-ITS 2023
Workshops: DriveX (@ECCV 2026), ROAD++ (@ICCV 2023)

PhD Jury Member:
Loick Chambon (Advisor, 2026)
Florent Bartoccioni (Invited Member, 2023)