This research introduces LLM-attacker, a novel framework that leverages large language models to enhance the effectiveness of adversarial testing for autonomous driving systems.

• Addresses the challenge of identifying which traffic participants to manipulate in testing scenarios
• Uses LLMs to prioritize potential adversarial agents based on context awareness
• Demonstrates superior performance in finding safety-critical edge cases compared to traditional methods
• Creates more realistic and diverse testing scenarios that better represent real-world risks

This innovation matters for engineering safer autonomous vehicles by significantly improving how we test their responses to dangerous situations, helping identify vulnerabilities before deployment.

LLM-attacker: Enhancing Closed-loop Adversarial Scenario Generation for Autonomous Driving with Large Language Models