Inherited, Not Invented: Tracing the Provenance of Wait in Reasoning Models

Inherited, Not Invented. Tracing the Provenance of Wait in Reasoning Models. By Zhizhuo Kou and eight others, from the Hong Kong University of Science and Technology, Peking University, and one other institution. Published on the twenty-second of October. Its code repository is worked out: about one hundred and fifty source files, with a readme and a licence.

Reinforcement learning is often credited with teaching reasoning models to write the word "wait" before they reconsider a step. Yet large accuracy gains from reinforcement learning do not actually require the word. The authors argue that the usual story confuses two things: what reinforcement learning discovers, and what a model inherits from its training data. Their claim is in the title. The marker is inherited, not invented.

To show it, they trace the word through the training histories of open models. They read it with label-free indices of frequency, position, form and usage, plus an injected-error test, all computed from sampled outputs. The first finding is that zero-RL almost never coins the marker. On base models that lack the word, reinforcement learning raises MATH-500 accuracy substantially, while "wait" stays close to absent. Instead, the word arrives with training data. In OLMo-3 it comes in during mid-training, and fine-tuning on reasoning traces installs it in most models whose base lacks it.

Where the base already uses the word, the picture depends on the recipe. Both released RL-Zero recipes, from OLMo-3 and MiMo, multiply it, while KL-free GRPO from the same base does not. The recipe of OLMo-3 also wears down the word's form, and turns it into a more sensitive but more liberal alarm. When reinforcement learning comes after supervised fine-tuning, it barely changes the marker at all. And the authors' own GRPO recipe erases a marker that fine-tuning had seeded.

Two cautions come out of this. Marker frequency alone does not indicate reasoning induced by reinforcement learning, so counting how often a model says "wait" is not a measure of what training taught it. And the authors caution against revision metrics that use an LLM as a judge, because automated judges can reverse trends that humans labelled. The evidence comes from open models whose training histories can be traced, and from the specific recipes named.

The shift here is one of attribution. A word that looked like a behaviour reinforcement learning had discovered turns out, in these histories, to be something the data put there, which training then amplifies, reshapes, leaves alone or removes depending on the recipe. Accuracy and the marker move separately, and the authors warn against reading one from the other.