I am an AI engineer and research student working on generative modeling, world models, reinforcement learning, and adaptive inference. My work is method-first: I study how learned dynamics, probability paths, numerical solvers, and inference-time computation affect prediction and decision-making 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. I use physical simulation, robotics, language agents, and financial markets as demanding testbeds for methods that can transfer across domains.
Submitted to the NeurIPS 2026 GDDL workshop. 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 finite 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. The result is a bounded non-implication: in the tested families, smoother average fields do not necessarily produce lower sampling error at a fixed number of function evaluations.
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.
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 from-scratch PyTorch implementation of Proximal Diffusion Models, with reverse-time SDE sampling, proximal-matching losses, low-NFE evaluation, tests, and reproducible Hydra configurations.
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.
- Generative AI Engineer, MBDA (apprenticeship, Sep 2025 - present): multi-agent LLM systems, vLLM, Docker, and GPU-cluster deployment.
- Data Scientist, TotalEnergies Digital Factory (apprenticeship, Nov 2023 - Sep 2025): machine learning for EV-charging pricing, public-station accessibility, and remote-sensing classification.
- 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 in an environment where I can develop general methods, publish rigorous work, and build systems at scale. My main interests are generative modeling, world models, adaptive inference, LLM agents, reinforcement learning, and scientific machine learning.
- LinkedIn: in/clement-callaert
- Email: clement.callaert@student-cs.fr
♟️ Outside research, I play chess at around 1750 Elo on chess.com.