Recent advances in reasoning models have significantly improved performance on complex tasks by enabling extended chains of thought, yet they also introduce a critical inefficiency: overthinking, where excessive reasoning yields diminishing or even negative returns. In this thesis, I investigate the behavior of reasoning models through the lens of overthinking and develop a structured framework for analyzing their reasoning dynamics.

I begin by identifying and characterizing the MIP-overthinking phenomenon, showing that reasoning models tend to generate unnecessarily long and ineffective reasoning traces, particularly under ill-posed settings such as missing-premise queries. This behavior reflects a fundamental limitation in current training paradigms, where models fail to appropriately allocate reasoning effort and exhibit weakened critical thinking abilities .

Building on this observation, I move beyond outcome-based evaluation and conduct a fine-grained analysis of reasoning traces. I introduce a structured perspective inspired by cognitive science—thinking episode theory—which decomposes reasoning into interpretable functional units. Using scalable automatic annotation, I analyze large-scale reasoning datasets and uncover consistent structural patterns governing reasoning behavior, including temporal dynamics, transition patterns, and their relationship to correctness.

Finally, I leverage this episode-level view to further examine how reasoning behaviors emerge, interact, and degrade under different settings. Our analysis reveals that reasoning performance is closely tied not only to the presence of specific reasoning components but also to their organization and transitions over time, providing new insights into both the strengths and limitations of current reasoning models.