I am an AI engineer and research student working on world models, generative modeling, reinforcement learning, and adaptive inference. I am particularly interested in when learned representations become useful for planning and decision-making, and in how inference-time computation should be allocated under limited budgets.
I am currently completing two degrees in parallel:
- an Engineering Degree at CentraleSupélec, majoring in Mathematics and Data Science;
- a Research Master's in Mathematics & Artificial Intelligence at Université Paris-Saclay.
Alongside my studies, I work as a Generative AI Engineer at MBDA, where I develop and evaluate multi-agent LLM systems and deploy self-hosted models on GPU infrastructure.
My projects emphasize controlled comparisons, reproducible artifacts, and explicit reporting of limitations and negative results.
A controlled comparison of factored and monolithic JEPA-style world models for planning on PushT, using a shared detector and MPPI planner. The apparent seed-0 planning gap, 12% versus 4%, disappeared at seed 1, 4% versus 4%. The factorization hypothesis was therefore not supported, and near-perfect in-slice probe scores did not reliably predict closed-loop planning performance.
An exact Gaussian Wasserstein study of whether averaged field regularity can rank probability paths for few-step generative sampling. I found 14 ranking inversions out of 36 comparisons across Euler, Heun, and RK4; the phenomenon replicated in 11 out of 18 comparisons in a preregistered non-centered family and passed an 80-digit precision audit. A manuscript based on this work is under review.
Research extensions to Latent Program Networks for adaptive test-time computation. On a held-out pattern_2d split, a REINFORCE stop-or-continue controller matched fixed-five exact match, 0.90625, while reducing mean search steps from 5.00 to 4.28. A gradient-norm heuristic achieved the same exact match in 3.44 steps, so the learned controller did not dominate the strongest simple baseline. These are controlled sandbox results, not ARC-AGI claims.
A reproducible offline-RL benchmark on FI-2010, built with Côme Genet. Across five seeds, the preregistered primary Decision Transformer did not beat buy-and-hold under the frozen cost model. The RTG sweep nevertheless changed deployed behavior substantially, moving mean normalized PnL from -1.976 at P10 to -0.033 at P90. Metrics, manifests, plots, and reproduction commands are committed.
An adaptation of ACES to search for regressive sycophancy in a local LLM using programmatic checkers. I diagnosed budget lock-in in a Bayesian ALP selector and introduced a fixed 20% uniform exploration floor. Across two seeds, mean archive coverage recovered from 0.792 to 0.917, while mean QD score increased from 9.219 to 11.000.
A from-scratch PyTorch implementation of Proximal Diffusion Models, with reverse-time SDE sampling, proximal-matching losses, low-NFE evaluation, tests, and reproducible Hydra configurations.
- Generative AI Engineer, MBDA (apprenticeship, Sep 2025 - present): multi-agent LLM systems, reinforcement learning for adversarial cybersecurity scenarios, vLLM, Docker, and GPU-cluster deployment.
- Data Scientist, TotalEnergies Digital Factory (apprenticeship, Nov 2023 - Sep 2025): machine learning for EV-charging pricing and public-station accessibility.
- M2 Mathematics & Artificial Intelligence, Université Paris-Saclay, 2025 - 2026
- Engineering Degree, Mathematics and Data Science, CentraleSupélec, 2023 - 2026
- Exchange in AI for Engineering, Beihang University, Beijing, 2025
- CPGE Mathematics and Physics, La Martinière Monplaisir, 2021 - 2023
I have also taught practical AI workshops with Automatants and led a six-person infrastructure team as president of Rézo Metz-Rennes Fédérés.
I am looking for a PhD or research-engineering position starting in autumn 2026, primarily in Europe, on world models, diffusion or flow models, adaptive inference, LLM agents, or reinforcement learning.
- LinkedIn: in/clement-callaert
- Email: clement.callaert@student-cs.fr
♟️ Outside research, I play chess at around 1750 Elo on chess.com.