Submitted by Songtao Wei 5 LEAD: Length-Efficient Adaptive and Dynamic Reasoning for Large Language Models University of Texas at Dallas 2 1