Standard LLM benchmarks assign each question a single outcome from one fixed wording. The paper points out that the same underlying request can be expressed in many equivalent ways, and this variability is hidden by typical benchmark scores.
The authors argue that the quantity of interest should therefore not be a single score. Instead, they propose accessibility profiles, which jointly measure the breadth and reliability of LLM capabilities.
This digest is based on the arXiv abstract; the full argument is not detailed in the source. The paper is available at the provided link.