Restricting a model's attention to only information directly relevant to the current reasoning step, removing access to unrelated context.