Reproduce “Auto-Tuning Safety Guardrails for Black-Box Large Language Models” on OpenResearch

About this paper

Perry Abdulkadir

A study demonstrates that treating external safety guardrails for black-box large language models as hyperparameters enables their efficient optimization using standard techniques like Optuna. This approach reduces the wall-clock time for identifying optimal guardrail configurations by approximately 8x compared to an exhaustive grid search, improving safety performance against adversarial prompts while controlling for helpfulness and latency.

Key insights

• External safety guardrails, comprised of system prompts and content filters, behave as tunable parameters for black-box LLMs, influencing their safety and helpfulness trade-offs.
• Standard black-box hyperparameter optimization (HPO) frameworks can efficiently identify high-performing guardrail configurations, significantly outperforming exhaustive grid search in terms of computational resources.
• Combining explicit safety system prompts with classifier-based content filtering offers a more robust defense than either mechanism alone, demonstrating efficacy in reducing malware generation and controlling over-refusal.

Summary by alphaXiv.