Standard Safety Benchmarks May Not Reliably Score Small Language Models

A large-scale empirical study evaluated five widely used AI safety benchmark suites against 26 open-source small language models (SLMs), scoring each response as harmful, safe, or ambiguous under a unified rubric. Across all five benchmarks, ambiguous judgments dominated the results, and the rate of ambiguity correlated with prompt complexity and model architecture rather than with actual safety behavior. The study found that ambiguity increases with lexical density, output perplexity, and output length, and decreases with lexical sophistication, self-coherence, and reply-prompt similarity, producing a capability-safety confound where model capability gets conflated with apparent safety. Because ambiguous responses are prevalent, aggregate mean-score leaderboards were shown to be mathematically brittle: model rankings shifted significantly depending on how ambiguous responses were treated in scoring, even when the underlying model outputs stayed the same. The paper is a research submission (arXiv, August 2025 listing) and does not indicate any fix or revised benchmark has yet been adopted.

rss · arXiv cs.AI · Aug 19, 04:00

「Background」 Organizations deploying small language models in resource-constrained or privacy-sensitive settings often rely on established LLM-centric safety, security, and compliance benchmark suites to demonstrate that a model meets safety expectations before or during deployment. These benchmarks were designed and validated primarily on large language models, and their scoring pipelines assume that model outputs can be cleanly classified as harmful or safe, an assumption this study tests directly against smaller models.

「Exposure」 This concerns any organization using standard, off-the-shelf safety or compliance benchmark suites to certify or justify the deployment of open-source small language models, particularly in on-device, edge, or privacy-sensitive contexts where SLMs are chosen over larger models. Teams should check which benchmark suites were used to produce safety claims for their deployed SLMs, whether those suites' scoring pipelines treat ambiguous or irrelevant responses as safe by default, and whether leaderboard rankings they relied on were sensitive to how ambiguous cases were counted. The finding spans 26 open-source SLMs and five widely used benchmark suites, so exposure is broad among adopters of small open models but the study did not test proprietary or closed SLMs.

「Mitigation」 No revised benchmark or automated fix is presented in the paper; the authors' main recommendation is that these benchmarks should not be used as standalone evidence of SLM safety. As a compensating control, organizations should treat aggregate leaderboard scores with caution, examine how ambiguous responses were classified in any benchmark results they cite, and supplement automated benchmarking with manual review or model-specific evaluation before relying on it for compliance purposes.

Tags: #benchmark validity, #small language models, #AI safety evaluation, #compliance, #model evaluation