Study reports fewer false refusals after paired-prompt training

Training a language model with paired harmful and benign requests sharply reduced mistaken refusals in a political-persuasion experiment, according to a preprint submitted to arXiv on September 3, 2026. The authors report that refusals of benign requests on held-out pairs fell from 32.94% to 4.16%.

The improvement came with a trade-off. Refusal of harmful requests in that comparison also fell, from 91.88% to 87.72%—a drop of 4.16 percentage points. The model became much more willing to answer the benign requests, but somewhat less reliable at rejecting the harmful ones.

The researchers tested Qwen3-8B on what they call “narrow-boundary safety”: teaching a model where to draw a line within a subject, rather than treating the whole subject as off-limits. Their example is politics. A deployment might need to reject targeted political manipulation while still answering factual questions about the same election. Harmful-benign pairs form part of their framework for both training and evaluation.

A separate result illustrates how refusal training can overshoot. Using refusal data completed through a step called Escalate, the authors raised refusal in the target domain from 9.47% to 84.75%. The mean unsafe-response rate across three broader harmfulness benchmarks fell from 26.26% to 0.14%. Yet over-refusal on XSTest rose from 2% to 74%.

Those figures are not one continuous before-and-after result. The 74% measures over-refusal on XSTest; the reduction from 32.94% to 4.16% comes from the authors’ held-out boundary pairs. The latter comparison supports a more specific finding: paired training improved discrimination in that test, without eliminating the trade-off between answering benign requests and refusing harmful ones. The reported results do not establish that the same gains carry over to other models or domains.