Ege C. Kaya

Ege C. Kaya

On the job market for industry research roles in reinforcement learning and LLM post-training

PhD Candidate in Electrical and Computer Engineering, Purdue University

I study reinforcement learning and stochastic optimization, with a focus on theory that explains an algorithm’s behavior. More recently, I’ve been working on efficient post-training methods for language models.

kayae [at] purdue [dot] edu CV Google Scholar LinkedIn GitHub

Research

I started with robust and submodular optimization, including work on sensor selection and distributed methods. My PhD research then moved toward reinforcement learning and stochastic approximation. I study distributional and average-reward RL, decision-making with coupled dynamics, and learning from dependent data or noise with unbounded variance.

More recently, I’ve been working on LLM post-training. I’m interested in when low-rank evolution strategies behave differently from dense Gaussian methods, and whether they can make training less expensive. Our EGGROLL and LOO-ROLL work looks at those questions through both analysis and experiments on language models.

Selected Papers and Preprints

  • Quotient-Categorical Representations for Bellman-Compatible Average-Reward Distributional Reinforcement Learning NeurIPS 2026, accepted Ege C. Kaya, Aliasghar Pourghani, Vijay Gupta, and Abolfazl Hashemi Advances in Neural Information Processing Systems, 2026. arXiv
  • Joint MDPs and Reinforcement Learning in Coupled-Dynamics Environments UAI 2026 Oral Ege C. Kaya, Mahsa Ghasemi, and Abolfazl Hashemi Uncertainty in Artificial Intelligence, 2026. Oral presentation. Top 2.2% of 1,087 submissions. arXiv
  • Stochastic Dominance Driven First-Order Policy Optimization for Multi-Objective Reinforcement Learning Ege C. Kaya, Kadierdan Kaheman, Jason M. Cloud, and Abolfazl Hashemi Uncertainty in Artificial Intelligence, 2026.
  • Randomized Greedy Methods for Weak Submodular Sensor Selection with Robustness Considerations Ege C. Kaya, Michael Hibbard, Takashi Tanaka, Ufuk Topcu, and Abolfazl Hashemi Automatica, 171, 2025.
  • Localized Distributional Robustness in Submodular Multi-Task Subset Selection Ege C. Kaya and Abolfazl Hashemi IEEE Transactions on Signal Processing, 2024.
  • EGGROLL, Unrolled: Understanding and Improving Low-Rank Evolution Strategies at Scale ICLR 2027, under review Ege C. Kaya and Abolfazl Hashemi arXiv
  • Why Does Adaptive Batching Help LLM Pretraining? A Perspective from Unbounded Variance ICLR 2027, under review Arda Fazla, Antesh Upadhyay, Ege C. Kaya, M. Berk Sahin, and Abolfazl Hashemi
  • Lower Bounds and Proximally Anchored SGD for Non-Convex Minimization Under Unbounded Variance ICLR 2027, under review Arda Fazla, Ege C. Kaya, Antesh Upadhyay, and Abolfazl Hashemi. First three authors contributed equally. arXiv
  • A Banach-Space Theory of Markovian Halpern Iteration for Non-Expansive Maps ICLR 2027, under review Ege C. Kaya, Arda Fazla, M. Berk Sahin, and Abolfazl Hashemi arXiv
  • A Finite-Iteration Theory for Asynchronous Categorical Distributional Temporal-Difference Learning AISTATS 2027, under review Ege C. Kaya and Abolfazl Hashemi arXiv
  • Learning to Control Coupled-Dynamics Environments with Joint Markov Decision Processes TMLR, under review Ege C. Kaya, Aliasghar Pourghani, Mahsa Ghasemi, Vijay Gupta, and Abolfazl Hashemi arXiv

Earlier Work

  • Equitable Client Selection in Federated Learning via Truncated Submodular Maximization. IEEE CDC, 2024.
  • Relative Entropy Regularization for Robust Submodular Multi-Task Subset Selection. Allerton, 2023.
  • Communication-Constrained Exchange of Zeroth-Order Information with Application to Collaborative Target Tracking. ICASSP, 2023.
  • High Probability Guarantees for Submodular Maximization via Boosted Stochastic Greedy. Asilomar, 2023.
  • Communication-Efficient Zeroth-Order Distributed Online Optimization: Algorithm, Theory, and Applications. IEEE Access, 2023.

Blog