I am an Assistant Professor at the Machine Learning department of the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI). My research focuses on areas like LLM reasoning, GFlowNets, uncertainty estimation, reinforcement learning sample complexity, and more broadly, on designing better "AI for science" tools.

Research Topics

AI for Science Tools

LLM Reasoning

Curriculum Learning

Uncertainty Estimation

Sample Efficient RL

Bayesian Optimization

GFlowNets

Probabilistic Modeling

My Contributions +

As AI systems are increasingly deployed in complex tasks and real-world applications, I believe it's crucial to understand and improve their reasoning and reliability. My work addresses challenges like quantifying uncertainty, performing efficient search and sampling, and ensuring reliable reasoning, all of which are fundamental for safe and effective AI deployment.

One major focus has been the development of Generative Flow Networks (GFlowNets), a framework that connects reinforcement learning and probabilistic modeling. I've worked on establishing the theoretical foundations of GFlowNets, demonstrating their effectiveness in Bayesian inference of discrete and continuous structures, and showing their ability to handle multimodal distributions better than traditional variational methods. I also created and released torchgfn, an open-source software framework, to make GFlowNets more accessible for both discrete and general applications.

Beyond GFlowNets, I've also contributed to uncertainty estimation through a method called DEUP (Direct Epistemic Uncertainty Prediction). DEUP addresses model misspecification in interactive learning settings by using a secondary predictor trained on the main model's errors. This results in more reliable uncertainty estimates for decision-making in downstream tasks.

Currently, I'm focused on scaling GFlowNets to tackle more complex problems, using them to enhance Large Language Model reasoning, and developing even better uncertainty estimation methods. My overall goal is to improve the reliability, efficiency, and reasoning capabilities of AI systems, particularly in scenarios that demand robust decision-making under uncertainty.

Not hiring in 2025! The usual form is closed.

News

• March 2025 - Our work "On the Privacy Risks of Spiking Neural Networks: A Membership Inference Analysis" has been accepted to the "I Can't Believe It's Not Better: Challenges in Applied Deep Learning" @ ICLR 2025 workshop as a spotlight talk.
• February 2025 - I gave a talk titled "GFlowNets: An Introduction and Recent Advances" at the MBZUAI-Berkeley Joint Workshop on Emerging Directions of Machine Learning.
• February 2025 - I gave a talk titled "GFlowNets: A Novel Framework for Diverse Generation in Combinatorial and Continuous Spaces" (slides) at the MBZUAI Paris Workshop 2025.
• February 2025 - Our work "PORT: Preference Optimization on Reasoning Traces" has been accepted to the NAACL 2025 conference.
• February 2025 - Our work "Customer Reactions to Companion AI: Exploring Opportunities and Threats for Vulnerable Consumers" has been accepted to the Frontiers in service 2025 conference.
• December 2024 - I gave a keynote talk titled "Advancing the Fourth Paradigm: Machine Learning Frameworks for Experimental Science" at the the MoroccoAI 2024 conference.
• November 2024 - I attended the CLIMB workshop, and the Domain adaption and related areas workshop at the Simon's institute, in University of California, Berkeley.
• October 2024 - I am co-organizing the NETYS 2025 conference in Rabat, Morocco, in May 2025.
• September 2024 - I joined MBZUAI as an Assistant Professor in the Machine Learning Department.

Ultimately, I am driven by a deep interest in understanding and defining intelligence, whether it's in animals or artificial systems. While I'm excited about the potential of AI, I'm also mindful of its limitations and potential pitfalls – I'm definitely not a technosolutionist.

Before my current role, I worked at TII as a Senior Researcher in 2024. I obtained my PhD in 2023 from Mila and UdeM, under the supervision of Yoshua Bengio. During my PhD, I also interned at Google Brain in Paris in 2020, focusing on hierarchical reinforcement learning. My academic background includes studies in applied mathematics at Ecole Polytechnique and statistical learning at Ecole Normale Supérieure Paris-Saclay. I've also explored research in game theory and operations research at IBM Research Singapore and gained industry experience as a data scientist at Booking.com in Amsterdam.