Benchmarking the Benchmarks: Evaluating Automated Safety Benchmarks for Small Language Models

Current safety benchmarks designed for large language models don't reliably work for smaller ones: The study found that existing AI safety benchmarks—tools used to test whether AI systems are safe—were built with large language models in mind and may not effectively evaluate smaller models that are increasingly deployed in real-world applications.

Ambiguous responses are a major problem that benchmarks can't handle: Across five widely-used benchmark suites tested on 26 different small language models, ambiguous or unclear responses dominated the results. The researchers note that "ambiguous judgments dominate," making it difficult to definitively say whether a model's response was actually safe or unsafe, which undermines confidence in benchmark scores.

Model capability and safety are being confused: The research reveals a "capability-safety confound"—meaning that a model's general intelligence level is being mixed up with its actual safety performance. Factors like how complex the output is or how well-written it is can make a model appear safer or less safe, even when the underlying safety hasn't actually changed.

Leaderboard rankings are unreliable: Because ambiguity is so prevalent in benchmark results, the rankings that compare different models' safety performance are mathematically unstable. The study found that "model rankings change significantly under reasonable ambiguity treatments," meaning small changes in how ambiguous responses are scored can completely flip which models appear safest.

We need better safety evaluation methods for smaller AI models: The paper concludes that "LLM-centric safety benchmarks are insufficient as standalone evidence for SLM safety assessment," suggesting that new evaluation approaches specifically designed for smaller models are needed before we can confidently deploy them in sensitive applications.