The wait marker in reasoning models was inherited from training data, not invented by RL

The marker nobody taught

Reasoning models have a tell: mid-chain they write a word like wait, then reconsider the step they just took. The standard story credits reinforcement learning with teaching them that habit. In a preprint posted this week, the authors trace the marker's provenance and find it was inherited, not invented. The marker arrives with training data: fine-tuning on reasoning traces installs it in most models whose base lacks it, while zero-RL models almost never coin it on their own. m:…5521

That flips the causal story. Released RL recipes like OLMo-3's and MiMo's multiply the marker, but only where the base already uses it — and the authors report that large RL accuracy gains do not require the marker at all. Marker frequency alone, they argue, does not indicate RL-induced reasoning. One more uncomfortable finding: their GRPO recipe erased a wait marker that fine-tuning had seeded. If you are judging a model's reasoning by its verbal tics, you are measuring the dataset. m:…5521

The dissociation holds across training recipes. On base models that lack the word entirely, RL still raises MATH-500 accuracy substantially while the marker stays near absent — reasoning improves, the tic does not appear. RL after supervised fine-tuning barely changes the marker, and KL-free GRPO from the same base does not multiply it. Even where the marker survives, the recipe distorts it: the OLMo-3 recipe wears down the marker's form, leaving a more sensitive but more liberal alarm. The marker is a readout of the data pipeline, not a gauge of the reasoning engine. m:…5521

Evals of deliberation may be fooled too

The measurement problem runs one layer deeper. The same preprint reports that automated judges can reverse human-labelled trends when scoring these markers — the judge's verdict can flip the direction of the human ranking. That is a warning for anyone running RL with AI feedback on reasoning traces: the reward signal may be grading the dataset's tics rather than the model's deliberation. m:…5521

This suggests a practical rule for builders: do not select reasoning checkpoints by marker frequency or by judge-scored thinking style. Measure the downstream task, and treat deliberation markers as data artifacts until proven otherwise. m:…5521

Deliberation is getting cheaper anyway

While one group dissects what thinking looks like, others are making it cost less. In another preprint posted this week, the authors of a diffusion-language-model study show that learned unmasking policies improve the speed-accuracy trade-off of masked diffusion LMs. A policy of just 856K parameters achieved the best accuracy-compute frontier on GSM8K, stayed competitive on MATH-500 and HumanEval, and transferred to Dream-7B-Instruct. The overhead math is kind: a 10% reduction in function evaluations offsets the policy's cost for policies up to 7.5M parameters. m:…9281

A third preprint takes the orthogonal route: stop reasoning early. Its BELT probe separates outcome prediction from stopping control, so one probe supports multiple accuracy-cost trade-offs through CPU policy refitting. On Qwen3-1.7B it cut generated tokens 41% and on OLMo-3-7B-Think 51%, without accuracy loss — and the probe transfers to coding and agentic tasks without ever being trained on code or agentic data. At 21.2x fewer parameters and 55x fewer training GPU-hours than OS-Pruner, it is a cheap way to buy back inference budget. m:…5975

The unmasking study also maps the policy's ingredients: hidden states are weak alone but can complement confidence signals in unmasking policies, and labeling predictions that match the final output is overly permissive at low function-evaluation budgets. Small learned policies win not by being clever but by being calibrated about what to reveal when. m:…9281

What changes for builders

The through-line is that reasoning's internals are becoming legible and tunable at inference time. The wait marker was a dataset artifact wearing an RL costume. Unmasking schedules in diffusion LMs can be learned rather than hand-set. And early stopping is now a reusable, transferable probe rather than a per-model hack. The implication for anyone serving reasoning models: the gains are increasingly at inference time, and they come from measuring the mechanism, not from bigger training runs. m:…5521