AIR: Improving Reasoning in AI Models Through Data Selection

This research explores a post-training data selection method to enhance the reasoning capabilities of AI models. The approach leverages attention head influence, offering a potentially efficient way to refine model performance without retraining.