Threshold testing

How it works

Many AI behaviours hinge on a threshold: refuse versus answer, flag versus allow, escalate versus handle. Threshold testing sweeps inputs from clearly-one-side to clearly-the-other and locates the transition, then asks whether it sits where policy intends. It surfaces both over-refusal (the boundary set too tight, blocking benign requests) and under-refusal (set too loose), and the unstable middle band where small changes flip the verdict.

When to use it

Tuning and auditing safety filters, content classifiers, and any allow/deny or confidence cutoff; diagnosing over- and under-refusal.

Limitations

Boundaries shift between versions and have to be re-mapped after upgrades, and a single threshold can hide very different behaviour across different request types.

Method yield

Findings

1

Versions spanned

4

Yield score

3

Severity-weighted across the published findings below. Why we measure this →

Findings it surfaces (1)

Documented failures this method catches — the evidence it works.

• Over-refusal of benign requestsMedium

  Safety tuning causes refusal of harmless requests that merely resemble sensitive ones.

  How it found it: Walking benign prompts across the refusal decision boundary maps exactly where harmless requests start getting blocked.

  Refusal

References & further reading

• XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models

  Röttger et al. · arXiv · August 1, 2023

Cite this