Does the word wait in model reasoning reveal scrutiny patterns?

The paper reports that models' chain-of-thought "thinking traces" are "predominantly (>98%) in English" regardless of the response language, with one exception: Mistral Small 4 119B, which "thinks within-language for Hindi, German, and Chinese." Within that thinking text, the authors find a sharp divergence in how often models generate the word "wait," a marker they treat as a sign of active scrutiny of the user's input. Qwen 3.5 models include "wait" in "over 95%" of their thinking traces, while "Mistral 4 Small" does so "at 17.6%," "Gemma 4 E4B at 5.9%," and "Gemini 2.5 Pro at 2.7%." The paper ties this gap to the pushback behavior it measures elsewhere in the excerpt: Qwen's high "wait" rate coincides with it showing the highest rates of pushback against false user statements among the models studied.

The authors frame the "wait" marker as evidence of "epistemic vigilance" — that Qwen "actively initiates scrutiny and fact checking of the user's input" during its reasoning, before it has decided how, or whether, to phrase a correction. This treats the chain-of-thought as a visible trace of whether a model is checking a claim at all, separate from the phrase-level ambiguity of its eventual response. The paper raises the pattern inside a section explicitly asking how to mitigate the broader pushback-ambiguity problem, which frames the "wait" rate as a candidate lever: a model that reliably scrutinizes claims in its reasoning may be a model that can be pushed toward less-hedged pushback in its output.

This is a mechanistic complement to Does hedged pushback from LLMs actually correct false user beliefs?, which measures how ambiguous a model's pushback is at the response level without explaining why some models push back more plainly than others; this note supplies one candidate internal correlate. It also bears on Do LLMs show reproducible psychological profiles when given standardized tests?, which finds reproducible model-specific behavioral signatures beneath a shared alignment-shaped pattern: a sharply model-specific "wait" rate, rather than a shared one, is a second axis on which models diverge even though the pushback behavior it feeds into is nominally governed by similar alignment training.

The excerpt does not establish causation — it says the "wait" rate "may explain" Qwen's pushback rate, not that it does — and gives only these four models' figures, not the full study's range. It also does not say whether the test-time mitigation described elsewhere in the paper uses this marker. What the excerpt supports is narrow but concrete: a simple lexical marker in a model's reasoning trace, countable automatically, already separates models by how much they appear to scrutinize a claim before responding, which makes it a cheap diagnostic worth checking even though the excerpt stops short of showing it can be used to intervene.