A paper under review by Kou and eight coauthors argues that reinforcement learning does not necessarily teach reasoning models to write “wait” before reconsidering a step. Across analyzed open models, the marker often appeared during mid-training or reasoning-trace fine-tuning, while reinforcement learning substantially improved MATH-500 accuracy without introducing it in models whose bases lacked the word. The authors also report that some reinforcement-learning recipes multiplied markers already present in the base, so the effect depends on the recipe. They further warn that marker frequency alone does not demonstrate reinforcement-learning-induced reasoning and that LLM-as-a-judge revision metrics may reverse human-labeled trends.