Marc Bellemare
Also known as Marc G. Bellemare, Marc Gendron-Bellemare
Marc Bellemare is a reinforcement-learning researcher whose work spans game benchmarks, agent learning, and real-world control. He helped develop the Arcade Learning Environment and coauthored work on distributional reinforcement learning, Dopamine, and Loon balloon navigation. A 2021 paper he coauthored received a NeurIPS Outstanding Paper Award; he later cofounded Reliant AI and worked on reinforcement learning for large language models. As of October 2026, he is vice president of modeling at Cohere in Montreal.
- Now at
- Cohere, McGill University and Université de Montréal
- Works
- 92
- Citations
- 36,844
- h-index
- 31
Publication numbers from OpenAlex, as of 15 October 2026.
Timeline
-
2012
The Arcade Learning Environment
He earns a PhD at the University of Alberta under Michael Bowling and Joel Veness. In 2012, he coauthors a paper with Yavar Naddaf, Veness, and Bowling that introduces the Arcade Learning Environment, a benchmark for general agents across Atari 2600 games.
-
2013–2018
DeepMind and Google Brain
He works at DeepMind in the United Kingdom from 2013 to 2017. In 2017, he coauthors a distributional-RL paper that models the full variation in rewards. In 2018, he is a Google Brain research scientist and coauthors Dopamine, a framework for deep reinforcement learning.
-
2020
Reinforcement learning for balloons
In 2020, he coauthors a Nature study on reinforcement-learning control of Loon superpressure balloons. The paper reports station-keeping deployments around the globe, including a controlled experiment over the Pacific that runs for 39 days.
-
2021
A NeurIPS Outstanding Paper Award
In 2021, “Deep Reinforcement Learning at the Edge of the Statistical Precipice,” a paper he coauthors with Rishabh Agarwal, Max Schwarzer, Pablo Samuel Castro, and Aaron Courville, receives a NeurIPS Outstanding Paper Award. It recommends bootstrap confidence intervals, performance profiles, and interquartile means for deep-RL comparisons.
-
2022
The Balloon Learning Environment
In 2022, Google Research thanks him as part of the Balloon Learning Environment team. The release provides an open-source, high-fidelity simulator and benchmark for deep-RL research on stratospheric balloon station-keeping.
-
2023
Co-founding Reliant AI
He co-founds Reliant AI with Karl Moritz Hermann and Richard Schlegel in 2023. The company builds a research workbench for biopharma teams to review literature and extract scientific and regulatory data.
-
2025
Reinforcement learning for language models
In 2025, he coauthors TOPR, a method for fine-tuning large language models with reinforcement learning. Its NeurIPS experiments cover the GSM8K and MATH reasoning benchmarks.
-
As of October 2026
Vice president at Cohere
As of October 2026, he is VP of Modeling at Cohere in Montreal and an adjunct professor at McGill University and Université de Montréal.
Sources and credits
Photo: We have no published photo of Marc Bellemare yet. Send us a free-licensed photo.
- Marc Gendron-Bellemare | Mila
- [1207.4708] The Arcade Learning Environment: An Evaluation Platform for General Agents
- Marc GENDRON-BELLEMARE - La recherche - Université de Montréal
- Going beyond average for reinforcement learning - Google DeepMind
- [1812.06110] Dopamine: A Research Framework for Deep Reinforcement Learning
- Introducing a New Framework for Flexible and Reproducible Reinforcement Learning Research
- Autonomous navigation of stratospheric balloons using reinforcement learning
- Announcing the NeurIPS 2021 Award Recipients - NeurIPS Blog
- The Balloon Learning Environment - Google Research
- Cohere acquires Reliant AI to expand sovereign enterprise AI for the global biopharma and healthcare sectors
- Tapered Off-Policy REINFORCE - Stable and efficient reinforcement learning for large language models