Entropy-Gradient Inversion: Moving Toward Internal Mechanism of Large Reasoning Models