"Wait" in Reasoning Models Turns Out to Be Inherited from Data, Not Discovered by RL

The word "wait" in chains of reasoning has become the calling card of reasoning models: the model writes it, stops, and returns to a previous step. The common view is that this is an invention of reinforcement learning, which found a useful habit of doubting on its own. A new paper from Hugging Face Daily Papers tested that claim against the training history of open models and reached a different conclusion. On base models where the word is absent, RL substantially raises accuracy on MATH-500, while "wait" stays nearly invisible. The marker arrives with the data.

How the researchers traced the origin of one word

The authors did not try to work out what the model "meant". They went through the training trajectories of open models and computed indices over samples of their responses, with no manual labeling. There are four indices: the frequency of the word, its position in the reasoning, its form (how exactly it is written and formatted), and the nature of its usage. The fifth tool is an injected-error test. A wrong step is planted in the reasoning, and the researchers check whether the model reacts to it with "wait". This shows whether the marker works as an alarm signal or just sits in the text for show.

The result for pure zero-RL is unambiguous: it almost never produces the marker. If the base model lacks the word, RL still raises MATH-500, but "wait" does not appear. Accuracy and this word are therefore less closely linked than is commonly assumed.

Where the word actually comes from

The marker arrives with the data. In OLMo-3, it is brought in by mid-training. In most models whose base does not use it, the word is installed by fine-tuning on reasoning traces.

The more interesting cases are those where the base model can already write "wait". Here the recipes diverge:

• both published RL-Zero recipes (OLMo-3 and MiMo) multiply the word's frequency;
• KL-free GRPO from the same base does not;
• the OLMo-3 recipe erases the marker's form and makes it a more sensitive but also more "liberal" alarm signal: it fires more often, including when there is no real error;
• RL after supervised fine-tuning barely changes the marker;
• the authors' own GRPO recipe removes a marker that fine-tuning had seeded.

The last point is especially telling. Depending on its settings, the same class of algorithms can amplify the word, leave it as is, or scrub it out. The behavior is determined by the specific recipe, not by the mere fact that RL was applied.

Pitfalls for anyone measuring reasoning

The main practical takeaway concerns metrics. The frequency of "wait" alone does not show that RL shaped the reasoning. A high count may mean the word was inherited from the data, and a low one does not mean the model cannot reconsider its solutions. On bases without the marker, RL produced gains without it.

The second warning is aimed at fans of automated evaluation. The authors urge more caution with revision metrics computed via LLM-as-a-judge: automated judges can reverse trends found by human annotators. If your pipeline computes the share of "self-checks" with another model, the numbers may show the opposite of what a human sees.

What to change in training practice

A team building its own reasoning model can take several lessons from this work.

• A rise in the frequency of "wait" in the logs cannot be counted as proof that RL taught the model to reason. First, find out whether the marker was present in the base, in mid-training, or in the SFT data.
• Recipes should be compared on the same base. The difference between RL-Zero in OLMo-3 and MiMo and KL-free GRPO shows that the outcome depends on the details.
• The injected-error test is cheaper and more reliable than counting words. It checks whether the model reacts to a real error, rather than merely writing the right word.
• Automated judges evaluating revisions need to be checked against human labeling on a sample, otherwise the trend can easily come out with the opposite sign.

For the industry, this shifts the conversation about where reasoning abilities come from. Part of the behavior that was attributed to RL is laid down by data at earlier stages. That matters for anyone deciding where to invest compute: in the quality of traces during mid-training and SFT, or in the RL loop itself.