Anas Barakat

Anas Barakat

Research Fellow

Singapore University of Technology and Design

Optimization · RL · Multi-Agent Learning

My research develops foundations and algorithms for sequential and strategic decision-making when objectives extend beyond expected cumulative reward, feedback is limited, and multiple adaptive agents interact. My work lies at the intersection of optimization, reinforcement learning, and learning in games. I am particularly motivated by emerging challenges in modern AI, including policy optimization for LLM post-training, preference-based learning, and decentralized learning in competitive and cooperative multi-agent systems.

I am currently a Research Fellow at Singapore University of Technology and Design, working with Georgios Piliouras and Antonios Varvitsiotis. Previously, I was a postdoctoral fellow at ETH Zurich working with Niao He. I received my PhD in mathematics and computer science from Institut Polytechnique de Paris (Télécom Paris), advised by Pascal Bianchi and Walid Hachem.

Optimization for Sequential and Strategic Decision-Making

My research develops foundations and algorithms for decision-making beyond classical assumptions on objectives, feedback, and agent interactions, spanning three complementary directions:

  • Policy optimization beyond expected additive rewards: Foundations and optimization algorithms for reinforcement learning with objectives beyond standard expected additive rewards, including general utilities and LLM post-training objectives.

  • Learning and optimization in structured games: Foundations and algorithms for learning in strategic and dynamic multi-agent environments, with a focus on the long-run behavior of learning dynamics and convergence to equilibria.

  • Online learning beyond convexity under limited feedback: Understanding when learning remains possible with nonconvex objectives and only partial or ordinal feedback. I develop algorithms and guarantees by exploiting latent structure—such as hidden convexity—and learning directly from bandit or preference feedback.

Selected & Recent Research

Learning in Continuous Games from Pairwise Preference Feedback
Anas Barakat
Preprint · arXiv
Why Pass@k Optimization Can Degrade Pass@1: Prompt Interference in LLM Post-training
Anas Barakat, Souradip Chakraborty, Khushbu Pahwa, Amrit Singh Bedi
NeurIPS 2026 · arXiv
Online Learning on Hidden-Convex Losses via Algorithmic Equivalence: Optimal Regret, Geometric Barrier, and Bandit Feedback
Anas Barakat, Andreas Kontogiannis, Vasilis Pollatos, Ioannis Panageas, Antonios Varvitsiotis
NeurIPS 2026 · arXiv
Online Learning in Stabilized Linear Dynamical Games with Adversarial Disturbances
Anas Barakat, John Lazarsfeld, Georgios Piliouras, Antonios Varvitsiotis
NeurIPS 2026 · arXiv
Convex Markov Games and Beyond: New Proof of Existence, Characterization and Learning Algorithms for Nash Equilibria
Anas Barakat, Ioannis Panageas, Antonios Varvitsiotis
AISTATS 2026 · Proceedings
On the Global Optimality of Policy Gradient Methods in General Utility Reinforcement Learning
Anas Barakat, Souradip Chakraborty, Peihong Yu, Pratap Tokekar, Amrit Singh Bedi
NeurIPS 2025 · Proceedings
Reinforcement Learning with General Utilities: Simpler Variance Reduction and Large State-Action Space
Anas Barakat, Ilyas Fatkhullin, Niao He
ICML 2023 · Proceedings

For a complete list of my publications, see below or my Google Scholar.

News

10/2026
Selected as a Top Reviewer at NeurIPS 2026.
10/2026
New preprint: Learning in Continuous Games from Pairwise Preference Feedback. arXiv
09/2026
Four papers accepted at NeurIPS 2026 on LLM post-training, online learning, and learning in games.
09/2026
Invited to serve as an Area Chair for ICLR 2027.
07/2026
Policy Gradients for Cumulative Prospect Theory in Reinforcement Learning accepted to TMLR.
05/2026
New preprint: Online Learning on Hidden-Convex Losses via Algorithmic Equivalence: Optimal Regret, Geometric Barrier, and Bandit Feedback. arXiv
05/2026
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
02/2026
Optimistic Online Learning in Symmetric Cone Games accepted to TMLR. Paper
01/2026
Convex Markov Games and Beyond accepted to AISTATS 2026. arXiv
10/2025
Designed and taught a new course with John Lazarsfeld and Iosif Sakos: 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

2024–present
Research Fellow
Singapore University of Technology and Design
Working with Antonios Varvitsiotis.
2022–2024
Postdoctoral Fellow
ETH Zürich, Department of Computer Science
Worked with Niao He.
2018–2021
Ph.D. in Applied Mathematics and Computer Science
Institut Polytechnique de Paris (Télécom Paris)
Advised by Pascal Bianchi and Walid Hachem.

Here is my CV for more information.

Research

Publications

Policy Optimization Beyond Expected Additive Rewards

Keywords: policy gradient methods, general utility RL, convex RL, LLM post-training.

Anas Barakat, Souradip Chakraborty, Khushbu Pahwa, Amrit Singh Bedi (2026). Why Pass@k Optimization Can Degrade Pass@1: Prompt Interference in LLM Post-training. NeurIPS 2026.

Arxiv

Olivier Lepel, Anas Barakat (2025). Policy Gradients for Cumulative Prospect Theory in Reinforcement Learning. Transactions on Machine Learning Research 2026.

Slides Journal Arxiv

Anas Barakat, Souradip Chakraborty, Peihong Yu, Pratap Tokekar, Amrit Singh Bedi (2025). On the Global Optimality of Policy Gradient Methods in General Utility Reinforcement Learning. NeurIPS 2025.

Poster Proceedings Arxiv

Kimon Protopapas, Anas Barakat (2024). Policy Mirror Descent with Lookahead. NeurIPS 2024.

Poster Proceedings Arxiv

Anas Barakat, Ilyas Fatkhullin, Niao He (2023). Reinforcement Learning with General Utilities: Simpler Variance Reduction and Large State-Action Space. ICML 2023.

Poster Slides Proceedings Arxiv

Ilyas Fatkhullin, Anas Barakat, Anastasia Kireeva, Niao He (2023). Stochastic Policy Gradient Methods: Improved Sample Complexity for Fisher-non-degenerate Policies. ICML 2023.

Poster Proceedings Arxiv

Anas Barakat, Pascal Bianchi, Julien Lehmann (2022). Analysis of a Target-Based Actor-Critic Algorithm with Linear Function Approximation. AISTATS 2022.

Poster Proceedings Arxiv

Learning in Structured Dynamical Games

Keywords: learning in games, game dynamics, multi-agent RL, Markov games with general utilities, Markov potential games, continuous games with state dynamics, game theory.

John Lazarsfeld, Anas Barakat, Georgios Piliouras, Antonios Varvitsiotis, Andre Wibisono (2026). When and Why is Optimistic Multiplicative Weights Slow? The Geometry of Energy Dissipation. NeurIPS 2026.

Arxiv

Anas Barakat, Ioannis Panageas, Antonios Varvitsiotis (2026). Convex Markov Games and Beyond: New Proof of Existence, Characterization and Learning Algorithms for Nash Equilibria. AISTATS 2026.

Arxiv

Anas Barakat, John Lazarsfeld, Georgios Piliouras, Antonios Varvitsiotis (2025). Online Learning in Stabilized Linear Dynamical Games with Adversarial Disturbances. NeurIPS 2026.

Slides Arxiv

Anas Barakat, Wayne Lin, John Lazarsfeld, Antonios Varvitsiotis (2025). Optimistic Online Learning in Symmetric Cone Games. Transactions on Machine Learning Research 2026.

Slides Journal Arxiv

Jiduan Wu, Anas Barakat, Ilyas Fatkhullin, Niao He (2025). Learning Zero-Sum Linear Quadratic Games with Improved Sample Complexity and Last Iterate Convergence. SIAM Journal on Control and Optimization 63 (5), 3244-3271.

Journal Arxiv

Pragnya Alatur, Anas Barakat*, Niao He (2024). Independent Policy Mirror Descent for Markov Potential Games: Scaling to Large Number of Players. IEEE CDC 2024. *corresponding author.

Slides Proceedings Arxiv

Philip Jordan, Anas Barakat, Niao He (2023). Independent Learning in Constrained Markov Potential Games. AISTATS 2024.

Poster Slides Proceedings Arxiv

Jiduan Wu, Anas Barakat, Ilyas Fatkhullin, Niao He (2023). Learning Zero-Sum Linear Quadratic Games with Improved Sample Complexity. IEEE CDC 2023.

Proceedings Arxiv

Online Learning under Limited Feedback

Keywords: online learning, bandit feedback, preference feedback, hidden convexity, non-convex optimization.

Anas Barakat (2026). Learning in Continuous Games from Pairwise Preference Feedback.

Arxiv

Anas Barakat, Andreas Kontogiannis, Vasilis Pollatos, Ioannis Panageas, Antonios Varvitsiotis (2026). Online Learning on Hidden-Convex Losses via Algorithmic Equivalence: Optimal Regret, Geometric Barrier, and Bandit Feedback. NeurIPS 2026.

Arxiv

Stochastic Optimization

Keywords: Stochastic approximation, Dynamical systems, non-convex stochastic optimization, adaptive gradient methods.

Anas Barakat, Pascal Bianchi, Walid Hachem, Sholom Schechtman (2021). Stochastic optimization with momentum: convergence, fluctuations, and traps avoidance. Electronic Journal of Statistics 15 (2), 3892-3947.

DOI Arxiv Journal

Anas Barakat, Pascal Bianchi (2020). Convergence and Dynamical Behavior of the ADAM Algorithm for Non-Convex Stochastic Optimization. SIAM Journal on Optimization, 31 (1), 244-274.

Poster Slides Video DOI Arxiv Journal

Anas Barakat, Pascal Bianchi (2019). Convergence Rates of a Momentum Algorithm with Bounded Adaptive Step Size for Non-Convex Optimization. ACML 2020.

Poster Slides Proceedings Arxiv

Talks

Invited talk - 5th Symposium on Machine Learning and Dynamical Systems
Feb 10, 2026 2:00 PM Kyoto University, Kyoto, Japan
Invited talk - ICCOPT 2025
Jul 24, 2025 3:00 PM University of Southern California, Los Angeles, USA
Invited talk - Learning Theory and Applications Workshop, NTU
Apr 29, 2025 4:00 PM College of Computing and Data Science, NTU, Singapore
Invited talk - Finance and RL Talks
Apr 21, 2025 10:00 PM Online
Invited talk - 4th Symposium on Machine Learning and Dynamical Systems
Jul 8, 2024 2:45 PM Fields Institute, Toronto, Canada
See all talks

Student Mentoring

  • Olivier Lepel — Master’s thesis, Mar.–Sep. 2024
    Policy Gradients for Cumulative Prospect Theory in Reinforcement Learning. TMLR 2026.
    Currently: Data Scientist at BCG.

  • Kimon Protopapas — Master’s semester project, Sep. 2023–Jan. 2024
    Policy Mirror Descent with Lookahead. NeurIPS 2024.
    Currently: Research Intern at University of Basel.

  • Philip Jordan — Master’s thesis, May–Dec. 2023
    Independent Learning in Constrained Markov Potential Games. AISTATS 2024.
    Currently: PhD student at EPFL.

  • Jiduan Wu — Master’s thesis, Oct. 2022–Mar. 2023
    Learning Zero-Sum Linear Quadratic Games with Improved Sample Complexity and Last-Iterate Convergence. IEEE Conference on Decision and Control (CDC) 2023, SIAM Journal on Control and Optimization, 2025.
    Currently: PhD student at Max Planck ETH Center for Learning Systems.

  • Harish Rajagopal — Master’s thesis, Mar.–Sep. 2022
    Multistage Step Size Scheduling for Minimax Problems.
    Currently: Quantitative Technologist at Qube Research & Technologies.

  • Julien Lehmann — Master’s thesis, May–Oct. 2021
    Analysis of a Target-Based Actor-Critic Algorithm with Linear Function Approximation. AISTATS 2022.
    Currently: Research Engineer at OCamlPro.

Teaching

SUTD (2024-2025):

ETH Zurich (2022-2024):

Télécom Paris (2018-2021): Teaching Assistant

Optimization for Machine Learning (graduate level), Statistics (graduate level), Probabilities (undergraduate level).

Academic Service

  • Area Chair: ICLR 2027.
  • Conference Reviewing: NeurIPS (‘22, ‘23, ‘25, Top Reviewer ‘26), ICML (‘25, Gold Reviewer ‘26), ICLR (‘25, ‘26), AISTATS (Top Reviewer ‘22, ‘24, ‘25, ‘26), EC (‘26).
  • Journal Reviewing: Journal of Machine Learning Research (JMLR), Transactions on Machine Learning Research (TMLR), Mathematical Programming, SIAM Journal on Optimization (SIOPT), Journal of Optimization Theory and Applications (JOTA), IEEE Transactions on Automatic Control.

Contact