HKUST Traces 'Wait' in Reasoning Models to Training Data, Not RL — 13 of 14 SFT Runs Inherit It

A new paper from HKUST (with Huawei Noah's Ark and Peking University) tracks the 'wait' revision marker across 120 open model endpoints and finds 17 of 18 zero-RL runs stayed below 0.05 'wait' per 1k tokens despite +42.9 median accuracy gain, meaning RL rarely coins the marker — 13 of 14 SFT runs on reasoning traces inherit it instead. RL recipes like OLMo-3 and MiMo amplify the marker and erode its form ('Wait,' → 'wait'), and the authors warn that LLM-judge behavioral assessments can reverse human-labeled trends.