LoCoT: Measuring Ineffective Chain-of-Thought Reasoning Tokens in Large Language Models

Chain-of-thought prompting improves large language model performance by externalizing intermediate reasoning steps, yet a fundamental question remains unanswered: at what point during generation has the model already committed to its final answer? We introduce the reasoning lock-in point, the earliest position in a chain-of-thought trace beyond which truncation no longer affects answer accuracy, as a principled lens for measuring reasoning trace efficiency. Using a controlled truncation protocol, we systematically characterize lock-in behavior across three model families and two mathematical reasoning benchmarks spanning five difficulty levels. Our study reveals that models commit to their answers substantially before trace termination, that lock-in depth is structured by problem difficulty and model capability in consistent and interpretable ways, and that the tokens generated after lock-in constitute a measurable and recoverable source of computational waste. These findings suggest that current chain-of-thought generation is systematically inefficient in a difficulty conditional manner, and that reasoning traces carry a meaningful distinction between tokens that determine the answer and tokens that merely elaborate it. We discuss implications for inference-time efficiency, difficulty-aware token budgeting, and the use of lock-in point as a diagnostic for reasoning quality in large language models.