Benchmark of benchmarks: analyzing the influence and code repository quality of LLM safety benchmarks

The researchers conduct a systematic measurement study of 31 LLM safety benchmarks (covering prompt injection, jailbreak, and hallucination), using 382 non-benchmark papers as a comparison, combining automated static analysis, a cumulative 220+ hours of manual runnability testing, and bibliometric analysis. The results show that only 39% of benchmark repositories run without modification, only 16% provide flawless installation guides, and only 6% state ethical considerations despite containing potentially harmful content, and these deficiencies show no significant improvement over the study period. A factor analysis finds that the degree to which a benchmark is adopted is related to author prominence and code runnability, but not to code quality standards such as Pylint scores or maintainability. The authors point out that some repositories publicly expose harmful content such as successful jailbreak responses without ethical warnings or access controls, effectively becoming undefended attack resources; the need for temporary modifications to run a benchmark may also make safety evaluations across different papers difficult to compare. On this basis, the researchers propose a checklist to help benchmark contributors improve code quality, documentation, and ethical practice.