Submitted on October 22, 2026, the paper examines the {wait} marker in open reasoning models and asks whether reinforcement learning creates it.
The authors report that zero-RL runs can substantially improve MATH-500 accuracy while {wait} remains nearly absent when the base model lacks the word. In OLMo-3, they locate its appearance during mid-training, before RL. Fine-tuning on reasoning traces installs the marker in most models whose bases do not use it. Released RL-Zero recipes for OLMo-3 and MiMo multiply an inherited marker, while KL-free GRPO from the same base does not.
The paper also reports that RL can change the marker into a more sensitive but more liberal error alarm, and that an LLM judge can reverse a human-labelled revision trend. It does not establish that these results apply to closed models, that any single reflection token measures reasoning quality, or that one training recipe explains all visible reasoning traces. The paper is under review.