SG-Bench: Evaluating LLM Safety Generalization Across Diverse Tasks and Prompt Types
This addresses the need for more comprehensive safety evaluation in LLMs, though it is incremental as it builds on existing benchmarks by adding tasks and prompt variations.
The authors tackled the problem of evaluating large language model (LLM) safety by developing SG-Bench, a benchmark that integrates generative and discriminative tasks and tests across various prompt types, revealing that most LLMs perform worse on discriminative tasks and are highly susceptible to prompts, indicating poor safety generalization.
Ensuring the safety of large language model (LLM) applications is essential for developing trustworthy artificial intelligence. Current LLM safety benchmarks have two limitations. First, they focus solely on either discriminative or generative evaluation paradigms while ignoring their interconnection. Second, they rely on standardized inputs, overlooking the effects of widespread prompting techniques, such as system prompts, few-shot demonstrations, and chain-of-thought prompting. To overcome these issues, we developed SG-Bench, a novel benchmark to assess the generalization of LLM safety across various tasks and prompt types. This benchmark integrates both generative and discriminative evaluation tasks and includes extended data to examine the impact of prompt engineering and jailbreak on LLM safety. Our assessment of 3 advanced proprietary LLMs and 10 open-source LLMs with the benchmark reveals that most LLMs perform worse on discriminative tasks than generative ones, and are highly susceptible to prompts, indicating poor generalization in safety alignment. We also explain these findings quantitatively and qualitatively to provide insights for future research.