I am Gaël (pron. [gah-el]), senior machine learning research scientist at LawZero, an AI safety nonprofit organization created by Yoshua Bengio and incubated at Mila - Quebec AI Institute. I work on the Scientist AI, an approach to building safer frontier models by disentangling predictive capability from agency. My interests include pre-training alignment, prospective monitoring, world modelling and scalable oversight.

Before joining LawZero, I completed my Ph.D. at the University of Auckland on deep causal modelling for reasoning and generalization. My background is in causal reasoning, which I see as crucial for creating safe and reliable frontier models that better understand the world and are more aligned as capability increases.

Feel free to reach out for research collaborations or discussions.

News

  • Jun 2026Co-authored a new paper on the Scientist AI: Safety from Honesty in a Disinterested AI Predictor.
  • Jun 2026Started mentoring for the Pivotal Research AI Safety Fellowship.
  • May 2026The Causal Cartographer was accepted at the 2nd Workshop on Compositional Learning @ ICML 2026.
  • Jan 2026Humanity’s Last Exam, a benchmark of expert-level questions to which I contributed causal inference questions, was published in Nature.
  • Jan 2026Moved with LawZero as a senior machine learning research scientist when it became an organization independent from Mila.
  • Sep 2025Completed my Ph.D. at the University of Auckland and joined Mila - Quebec AI Institute to work on the Scientist AI as a senior machine learning research scientist.

Experience & Education

  • 2026-presentSenior Machine Learning Research Scientist, LawZero, Montréal.
  • 2026Research mentor, Pivotal Research Fellowship.
  • 2025Senior Machine Learning Research Scientist, Mila - Quebec AI Institute, Montréal. Scientist AI project, which I have continued at LawZero since it became independent of Mila in 2026.
  • 2024-2025Research Assistant, University of Auckland. Led the research and development of neural-causal networks for behaviour discovery in multi-agent systems.
  • 2021-2025Ph.D. in machine learning, University of Auckland. Thesis: A Causal Modelling Approach to Reasoning and Generalisation in Deep Neural Networks, supervised by Prof. Michael Witbrock and Prof. Gillian Dobbie.
  • 2023-2024Ph.D. mentor, University of Auckland. Mentored a first-year Ph.D. student on thesis scoping, problem definition and academic writing.
  • 2022-2023Graduate Teaching Assistant, University of Auckland. Advanced Machine Learning (M.Sc.) and Complexity, Algorithms and Graph Theory (B.Sc.).
  • 2021-2022Software Engineer, Alten / Amadeus IT Group (contractor). C++ development for a key component of Amadeus’ Global Distribution System, enhancing scalability of the retrieval pipeline.
  • 2020Dipl. Ing. (eq. M.Sc.) in software engineering, National Institute of Applied Sciences (INSA) of Rennes, and dual M.Sc. in machine learning, University of Rennes 1.

Reviewing: NeurIPS (top reviewer in 2025 and 2026), ICML, ICLR, AAAI (program committee member), COLM, EMNLP and ECML PKDD.

Awards & Talks

Research

During my Ph.D., I evaluated large language models (LLMs) and pointed out their brittleness when learning abstract representations. I showed that inducing LLMs to behave as causal models can improve the learning of robust and domain-invariant mechanisms. I also worked on extracting causal structures from natural language with LLM agents and harnessing these structures to make the model reason more causally from less information. At LawZero, I build on this to create models that can make predictions about the world without pursuing goals of their own.

Selected publications

  • Safety from Honesty in a Disinterested AI Predictor
    Yoshua Bengio, Oliver Richardson, Tomáš Gavenčiak, Michael Cohen, Rory Svarc, Damiano Fornasiere, Gaël Gendron, David Hyland, Aton Kamanda, Adam Oberman, Francis Rhys Ward, Anna Gavenčiak, Jacob Livingston Slosser, Vincent Mai, Iulian Serban, Joumana Ghosn
    Preprint 2026

    A formal safety argument for the Scientist AI: a predictor trained to approximate a Bayesian posterior, never rewarded for the consequences of its predictions, is unlikely, under stated assumptions, to cause harm through coordinated deception.

  • The Scientist AI: Safe by Design, by Not Desiring
    Damiano Fornasiere, Oliver Richardson, Gaël Gendron, Iulian Serban, Yoshua Bengio
    Preprint 2026

    A design for AI that is highly capable yet holds no goals of its own by construction: limit affordances and goal-directedness, attribute claims to their sources, and withhold feedback on downstream outcomes.

  • Causal Cartographer: From Mapping to Reasoning Over Counterfactual Worlds
    Gaël Gendron, Jože M. Rožanec, Michael Witbrock, Gillian Dobbie
    ICML 2026 Workshop · 2nd Workshop on Compositional Learning @ ICML 2026

    LLM agents that extract causal knowledge from text into a causal graph (CausalWorld), then reason step by step over counterfactual worlds, with competitive accuracy and up to 70% lower inference cost than chain-of-thought.

  • Can Large Language Models Learn Independent Causal Mechanisms?
    Gaël Gendron, Bao Trung Nguyen, Alex Yuxuan Peng, Michael Witbrock, Gillian Dobbie
    EMNLP 2024 · Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

    Independent Causal Language Models (ICLM): a modular architecture of independent causal modules selected by a router, improving out-of-distribution reasoning with accuracy gains of up to 40 percentage points over LoRA fine-tuning on the evaluated tasks.

  • Large Language Models Are Not Strong Abstract Reasoners
    Gaël Gendron, Qiming Bao, Michael Witbrock, Gillian Dobbie
    IJCAI 2024 · Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence

    A benchmark of text-based abstract reasoning tasks showing that LLMs, whether prompted or fine-tuned, struggle to generalize to unseen compositions of abstract rules.

    Also presented at the How Far Are We From AGI Workshop @ ICLR 2024
  • Disentanglement of Latent Representations via Causal Interventions
    Gaël Gendron, Michael Witbrock, Gillian Dobbie
    IJCAI 2023 · Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence

    CT-VAE: a variational auto-encoder that disentangles latent factors through hard causal interventions on a quantized latent space.

    University of Auckland Best Student Published Paper in Computer Science (2023)

Other publications

I also contributed causal inference questions to the Humanity’s Last Exam benchmark, published in Nature (2026), and abstract reasoning tasks to OpenAI’s Evals.

See the complete list on Google Scholar.