Can LLMs learn to think about their own thinking?

Existing methods reward language models either for final answers or by hand-crafting task-specific quality rubrics—both limiting. This work introduces MaR, which treats reasoning itself as trainable through metacognition: models learn to identify task-relevant information and adjust their problem-solving strategy mid-process, not just chase correct answers. Tested on 22 benchmarks, MaR lifts Qwen3.5-9B to compete with much larger proprietary models and generalizes to unseen tasks.