Research papers · 2025.11.18
From Words to Collisions: LLM-Guided Evaluation and Adversarial Generation of Safety-Critical Driving Scenarios
From Words to Collisions: LLM-Guided Evaluation and Adversarial Generation of Safety-Critical Driving Scenarios
Overview
The approach combines structured scenario parsing with Cartesian and ego-centric prompts for LLM safety evaluation. An adversarial module modifies threatening trajectories, with multiple models evaluated in a two-dimensional simulation environment.
Author-supplied source information
Author-supplied information on arXiv.
Final Version and Paper Accepted at IEEE ITSC 2025
Version: arXiv v4 — manuscript revision