Anas Barakat
Postdoctoral Research Fellow
Singapore University of Technology and Design
Learning, Optimization, and Game Dynamics for Modern AI Systems
Theory and algorithms for policy optimization beyond expected additive rewards, multi-agent learning in structured games, and online learning under limited feedback
My research develops theoretical foundations and algorithms for learning problems that go beyond the classical paradigm of maximizing expected additive rewards in fixed learning environments. In modern AI systems, objectives may be nonlinear, distributional, multi-sample, or nonconvex with exploitable hidden structure; learning feedback may be limited and revealed online; and environments may contain other agents adapting strategically. These challenges arise in modern AI training and LLM post-training, where models are optimized through imperfect objectives, learned rewards, verifiers, and interactions with users or other agents.
My research is organized along three complementary axes:
-
Policy Optimization Beyond Expected Additive Rewards: Theory and algorithms for reinforcement learning and post-training objectives beyond standard expected additive rewards, including general utilities, risk-sensitive criteria, and LLM post-training.
-
Multi-Agent Learning in Structured Dynamical Games: Learning dynamics in strategic, stateful multi-agent environments with exploitable structure, including Markov games with general utilities, Markov potential games, symmetric cone games and continuous dynamical games.
-
Online Learning with Hidden Structure under Limited Feedback: Online learning with hidden-convex losses, bandit feedback, and dynamically evolving state-dependent losses with adversarial disturbances.
Selected Publications
Policy Optimization Beyond Expected Additive Rewards
Keywords: general utility RL, policy gradient methods, risk-sensitive RL, LLM post-training.
Multi-Agent Learning in Structured Dynamical Games
Keywords: multi-agent RL, Markov games with general utilities, Markov potential games, game dynamics, Nash equilibrium learning, hidden convexity, symmetric cone games, dynamical games.
Online Learning with Hidden Structure under Limited Feedback
Keywords: online learning, bandit feedback, hidden convexity, regret analysis.
For a complete list of publications, see my Research page or Google Scholar.
News
- 05/2026: New preprint: Online Learning on Hidden-Convex Losses via Algorithmic Equivalence: Optimal Regret, Geometric Barrier, and Bandit Feedback, arxiv link.
- 05/2026: Humbled to be recognized as a Gold Reviewer at ICML 2026.
- 02/2026: New preprint: Why Pass@k Optimization Can Degrade Pass@1: Prompt Interference in LLM Post-training, arxiv link.
- 02/2026: Optimistic Online Learning in Symmetric Cone Games accepted to Transactions on Machine Learning Research, link.
- 01/2026: Convex Markov Games and Beyond accepted to AISTATS 2026, arxiv link.
- 10/2025: Designing and teaching a new course with John Lazarsfeld and Iosif Sakos. Check out our website: Online Learning and Learning in Games.
- 09/2025: On the Global Optimality of Policy Gradient Methods in General Utility Reinforcement Learning accepted to NeurIPS 2025.
Experience and Background
- October 2024–present: Postdoctoral Research Fellow, Singapore University of Technology and Design. Working with Antonios Varvitsiotis.
-
February 2022–August 2024: Postdoctoral Fellow, ETH Zurich, Department of Computer Science. Worked with Niao He.
-
Ph.D. in Applied Mathematics and Computer Science, Institut Polytechnique de Paris (Télécom Paris), 2021.
Advised by Pascal Bianchi and Walid Hachem. -
Engineering Master’s degree in Applied Mathematics and Computer Science, Télécom Paris, 2018.
-
M.Sc. in Data Science, Université Paris Saclay, 2018.
Here is my CV for more information.