Tag
BPCO introduces a stable recipe for critic-based reinforcement learning in language models, combining bounded value predictions and Monte Carlo targets to match group-based methods with single-response sampling.