Mechanistic Unveiling of Transformer Circuits: Self-Influence as a Key to Model Reasoning