INVESTIGATING REASONING MODELS: OVERTHINKING AND THINKING EPISODE THEORY

Abstract

Large language models (LLMs) have made remarkable progress in complex reasoning tasks by generating explicit chains of thought. However, the quality, efficiency, and structure of these reasoning processes remain poorly understood. This thesis investigates reasoning models through three complementary studies.

First, we identify a critical failure mode called Missing Premise Overthinking (MiP-Overthinking), where reasoning models generate excessively long responses to ill-posed questions lacking necessary premises. We show that this phenomenon contradicts the test-time scaling law: more computation does not help models identify the missing information. Surprisingly, non-reasoning models demonstrate better critical thinking ability in these scenarios.

Second, we introduce a novel analytical framework grounded in Schoenfeld's Episode Theory, a cognitive science framework originally developed for understanding human mathematical problem-solving. By annotating reasoning traces with cognitive episode labels such as Read, Analyze, Plan, Implement, Explore, Verify, and Monitor, we provide the first principled methodology for dissecting the reasoning process of large reasoning models.

Third, we scale this episode-level analysis into ThinkARM (Anatomy of Reasoning in Models), a framework that enables automatic, large-scale annotation and comparison of reasoning traces across 15 diverse models. Our analysis reveals consistent temporal organization in reasoning traces, identifies structural differences between reasoning and non-reasoning models, and demonstrates that episode-level patterns correlate with solution correctness and the behavioral impact of efficient reasoning methods.

Together, these three studies advance our understanding of how reasoning models think, where they fail, and how their reasoning processes can be systematically characterized.