Prompt-Induced Score Variance in Zero-Shot Binary Vision-Language Safety Classification
Factual evidence
What the source reports
Research demonstrates first-token AI safety scores vary widely across minor prompt rewordings, introducing reliability risks in guardrails.
Inspect the evidence
- Inclusion basis
- Enterprise AI
- Publisher and source type
- arXiv cs.CL — Computation and Language · RESEARCH
- Published by source
- 27 October 2026
- Collected by OneBench
- 28 Oct 2026, 02:01 UK
Stored source excerpt
arXiv:2605.00326v2 Announce Type: replace Abstract: Safety pipelines often block content by thresholding a model's first-token unsafe score. We show that a single prompt's score is…
Short excerpt from the collected text, not the full source. Use the source link to read it in context.
The factual summary is a OneBench synthesis, not a quotation or independent verification. Collection time is not publication time. Open the source for its full context; related reporting can share the same underlying announcement.
OneBench interpretation
Institutional assessment
So what
Single-prompt thresholding for AI safety checks introduces arbitrary score variance that undermines automated content and transaction moderation controls.
Do what
Review safety classifier prompt sensitivity and calibration methods with the team responsible for AI model risk and governance.