Can simple discourse markers make LLMs question harmful assumptions?

Cheng, Hawkins and Jurafsky report that two minimal prompting interventions produce large gains on three safety benchmarks previously described as resistant to mitigation: directly instructing the model to correct false presuppositions or misconceptions ("EXPLICIT"), and prepending the discourse marker "wait a minute" to the model's output ("WAIT"). On Cancer-Myth, where Zhu et al. (2026) "report that no LLM gets a score higher than 0.30," the paper's interventions "achieve score 0.51 (70% increase) while also controlling for false positives." Across six state-of-the-art LLMs, the authors report improvements nearly 4x for Cancer-Myth and a 40% increase on SAGE-Eval... across nearly all dimensions of the three benchmarks, while minimizing false positive rates (excessive challenging).

The explanation offered is pragmatic rather than informational: "wait a minute" is proposed to work not because it adds facts but because, as a discourse marker, it signals that the upcoming utterance is about to reconsider or interrupt the preceding common ground — the same cue that would prompt a human listener to pause accommodation and invoke epistemic vigilance. The excerpt notes this sits alongside a separately documented effect where adding "wait" improves reasoning performance by allocating more test-time compute (citing Muennighoff et al. 2025 among others), a mechanism the authors distinguish from their own pragmatic-cue account, though the excerpt does not disentangle how much of the Cancer-Myth and SAGE-Eval gain is pragmatic-cue-driven versus compute-driven.

This is the applied counterpart to Why do language models fail to challenge harmful user beliefs?, which supplies the pragmatic explanation these interventions are designed against: if accommodation is the default and vigilance is gated by at-issueness, assertion, and source reliability, a cue like "wait a minute" works by artificially triggering the same linguistic conditions that would make a human listener challenge a claim. It also complicates Why do language models avoid correcting false user claims?, which frames grounding failure as resistant to anything short of a training-level fix — these results suggest a minimal inference-time discourse marker can shift that strategy without retraining, at least on these three benchmarks.

The excerpt tests only two interventions, in English only, and flags that over-applying them "may risk over-correcting the user and diminishing user experience" — it does not report how often WAIT or EXPLICIT produce an unwanted gotcha tone in benign conversation, nor whether gains generalize beyond these three constructed benchmarks. The authors also leave open how much of the "wait" effect is their proposed pragmatic-cue mechanism versus the separately documented test-time-compute effect of the same token on reasoning tasks. At the strength the evidence supports, the implication is that some LLM safety failures framed as requiring retraining or fine-tuning may instead be addressable with small, interpretable, inference-time pragmatic cues — a narrower and more testable claim than models need better alignment.