We propose Learning to Focus (LeaF), which identifies and masks confounding tokens via gradient‐based comparisons, thereby improving long‐context reasoning accuracy and interpretability in large language models.