Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models
The study investigates how large language models form answers during Chain-of-Thought reasoning by identifying a 'commitment boundary.' It finds that most models transition from transient guesses to a stable final answer in a single step. Steps after this boundary are deemed 'epiphenomenal,' meaning they don't change the outcome. This allows for efficient reasoning truncation by exiting at the commitment point.