#EMA baseline

'NVIDIA's Reinforcement Learning Pretraining (RLP) injects dense, position-wise reinforcement into pretraining by rewarding chains-of-thought for information gain, improving reasoning and data efficiency across architectures.'