Researchers have introduced a reinforcement learning (RL) framework designed to enhance the reasoning capabilities of large language models (LLMs). This innovative approach aims to fine-tune LLMs, resulting in improved performance beyond traditional next-token prediction. The study highlights that the majority of training compute is now directed towards more complex training objectives, which allow models to learn intricate statistical features. Additionally, a novel method called multiagent finetuning has been proposed, enabling self-improvement through diverse reasoning chains. This method contrasts with conventional single-agent finetuning, which often experiences performance plateaus after limited iterations. The research involves contributions from institutions such as MIT CSAIL and Harvard University.