Will using RL further improve performance?

[FutureForMe]

This paper has been very inspiring to us, thanks. But I noticed that LIMO used 817 data to full-parameter fine-tune Qwen2.5-32B-Instruct. I would like to ask the author if there has been further research on the impact of Reinforcement Learning (RL) on the model’s performance. If so, do you think this approach could further enhance the model’s performance, similar to the improvements seen with DeepSeek-R1?