A new reinforcement learning framework called DiRL aims to enhance reasoning capabilities in large language models by distinguishing between genuine reasoning and memorization during exploration. The approach shows promising results in improving performance on reasoning tasks compared to existing methods.