PKPO

Pass@K Policy Optimization: Solving Harder Reinforcement Learning Problems

Christian Walder and Deep Karkhanis · NeurIPS 2025 Spotlight

Read paper · arXiv:2505.15201v5

Transform rewards to optimize a pass@k policy-gradient objective.

What you will implement

Continuous-reward transformation, policy loss, and the paper’s one-dimensional Gaussian toy.

• 45. PKPO Reward Transform and Unbiased Policy Gradient

  Implement expected maximum and the sloo-minus-one policy gradient.

  Supplied actions and log probabilities; continuous rewards and an explicit k=1 convention, without LLM training.