The ICRL framework enhances large language models by enabling them to internalize self-critique through reinforcement learning, improving their performance without reliance on external feedback. This approach shows significant gains in both agentic and mathematical reasoning tasks, suggesting a promising advancement in AI training methodologies.