EVOREFUSE: Evolutionary Prompt Optimization for Evaluation and Mitigation of LLM Over-Refusal to Pseudo-Malicious Instructions

We propose a novel bilevel system prompt optimization problem and a novel meta-learning framework to tackle it.