I recently wrapped up my Postdoctoral Fellowship at Mila – Quebec AI Institute, co-supervised by Yoshua Bengio and David Rolnick.

I study what training objectives actually install in a world model's latent state—objective→representation causality. My research centers on graph-structured world models and model-based reinforcement learning, with an emphasis on generalization across environments, network layouts, and dynamics. This grew out of applied work using RL to control real physical systems—building energy, where ~37% of global CO₂ lives—which taught me that the gap between predicting a system and controlling it is where the interesting questions are.

My work sits at the intersection of model-based RL, graph neural networks, and the physical world. I'm particularly interested in learning generalizable physics, and in measuring what training objectives actually install—across architectures.

• World Models and Intervention Operators — What does it mean for a latent world model to be good for control? Planning-time diagnostics for latent world models, and when value-equivalence is (and isn't) enough.

  • Operator-on-F Complements Value-Equivalence: A Planning-Time Diagnostic for Latent World Models — [accepted, RLC 2026 Workshop] on Model-Based RL in the Era of Generative World Models
  • The Rank-One Corner: How Much Value Equivalence Does a Task Need from a World Model? — sole-authored [preprint on arxiv]

• Graph Dreamer — A world model architecture for variable-size, heterogeneous graph environments that learns the structural relationships governing thermal dynamics, enabling zero-shot transfer across buildings with different topologies. (code temporarily private)

  • Graph Dreamer: Temporal Graph World Models for Sample-Efficient and Generalisable RL — WiML @ NeurIPS 2025 [OpenReview]
  • HVAC-GRACE: Transferable Building Control via Heterogeneous Graph Neural Network Policies — [ICML 2025 CO-BUILD Workshop] (94–97% transfer across building configurations)
  • Graphs for Scalable Building Decarbonisation: A Transferable Approach to HVAC Control — [NeurIPS 2025 Workshop] on Climate Change AI

• HOT Dataset — ~150,000 simulated controllable buildings and a benchmark gym for transfer-learning research.

  • A HOT Dataset: 150,000 Buildings for HVAC Operations Transfer Research — BuildSys 2025 [HuggingFace]

• In-Context & Bayesian RL — Bayesian fusion of context and value priors for in-context RL from suboptimal data.

  • Bayesian Decision-Time Inference for In-Context Reinforcement Learning from Suboptimal Data — accepted, RLC 2026 Workshop on Continual Reinforcement Learning [arXiv]

• Relational Models for MBRL — Empirical-mechanistic research on when and whether relational structure in world models helps.

  • When structural priors help prediction but hurt control" (in progress; code temporarily private)

📚 Full publication list on Google Scholar

Co-organizing the NeurIPS 2026 Smart Buildings Challenge: Learning Foundation Models for Real-World Optimization at Scale. Previously co-organized CoBuild @ ICML and UrbanAI @ NeurIPS (2023, 2024).

Two-time Olympian in modern pentathlon (London 2012, Rio 2016). Currently based in Nice, France. Board member at Racing to Zero, a nonprofit focused on sustainability in sport.