<ul><li>Pass-at-k Policy Optimization (PKPO) proposed as a fix for limitations in Reinforcement Learning algorithms optimizing for pass@1 performance.</li><li>PKPO transforms final rewards to optimize for pass@k performance, prioritizing sets of samples that maximize reward when considered jointly.</li><li>Novel low variance unbiased estimators derived for pass@k and its gradient in both binary and continuous reward settings.</li><li>PKPO enables robust optimization of pass@k for any arbitrary k <= n, allowing for annealing k during training to optimize both pass@1 and pass@k metrics.</li></ul>

Pass@K Policy Optimization: Solving Harder Reinforcement Learning Problems

Discover more