New framework audits safety labels in multimodal LLM benchmarks

Researchers have developed a new framework called the Benchmark Card to analyze the reliability of situational-safety benchmarks used for multimodal large language models (MLLMs). This framework identifies seven distinct fields that contribute to a benchmark's safety label, helping to distinguish between signals from the scene itself, the user's request, or the benchmark's internal files. The study found that some benchmarks, like MSSBench, inadvertently leak safety information through image files, while others, like MOSSBench and HoliSafe, rely on design elements or request phrasing rather than inherent scene safety. The research also introduced the Spot-Stop decomposition to better understand how MLLMs rank safety pairs, revealing that many models perform better at ranking than labeling. AI

IMPACT This research could lead to more reliable safety evaluations for multimodal LLMs, improving their deployment and trustworthiness.

RANK_REASON The cluster contains an academic paper detailing a new methodology for evaluating AI benchmarks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.MA (Multiagent) →

• Benchmark Card
• HoliSafe
• MOSSBench
• MSSBench
• multimodal large language models
• Spot-Stop decomposition
• VSCBench

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →