This research introduces LLM-attacker, a novel approach that leverages large language models to intelligently generate challenging test scenarios for autonomous driving systems.

• Uses LLMs to identify potential adversarial participants in traffic scenarios
• Implements a closed-loop framework that dynamically adapts testing based on vehicle responses
• Demonstrates superior effectiveness in generating safety-critical events compared to existing methods
• Shows practical applications for improving robustness of autonomous driving systems

This innovation matters because it addresses a critical engineering challenge: creating realistic yet challenging scenarios to test autonomous vehicles before real-world deployment, potentially saving lives by identifying safety vulnerabilities earlier in development.

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