PathCal: State-Aware Reflection-Marker Calibration for Efficient Reasoning

A new paper on arXiv introduces PathCal, a method for calibrating reflection markers in Large Reasoning Language Models (LRMs). LRMs generate long Chain-of-Thought trajectories containing markers like "wait", "but", and "alternatively", which signal hesitation, revision, or alternative exploration. Prior work treated these markers as a single category, but PathCal distinguishes their functional roles and temporal influence. Through type-wise suppression and fixed-prefix intervention, the authors show that different marker classes affect accuracy and generation length distinctly. The approach aims to improve reasoning efficiency by selectively controlling these markers during inference.