“The performance ceiling for models trained with reinforcement learning (RL) is higher than for those trained with supervised fine-tuning (SFT), even when the SFT data is high-quality human data.”