Bahram Behzadian

Research Scientist, Meta · Google Scholar

About

I work on reinforcement learning and sequential decision-making. The through-line across my research is policy improvement when the model of the world or the evaluator is imperfect: robust MDPs when the dynamics are uncertain, and truncated or approximate evaluators when the reward signal is unreliable.

My recent work studies policy improvement under verifier-style rewards, which connects directly to post-training and RL for reasoning models. Earlier work covered robust MDPs, learning under uncertainty, and partial observability, including scalable robust RL. I’m increasingly focused on model-based RL and reasoning systems that plan under imperfect models.

Research Interests

  • Policy improvement under truncated or approximate evaluators; verifier-style rewards and RLVR
  • Post-training and RL for reasoning and long-horizon decision problems
  • Robust RL and learning under uncertainty (robust MDPs, ambiguity sets)
  • Model-based RL, planning under learned or imperfect models
  • Evaluation methodology and stable learning dynamics

Contact

Email: behzadian.bahram@gmail.com