Safety monitoringExpert
Localized Conformal Safety Monitoring with Vision-Language Models for Autonomous Driving
This preprint introduces SLLCP, a post-hoc calibration layer over frozen vision-language models that turns raw safe/unsafe guesses into probabilistic safety prediction sets for autonomous driving. On 15,321 unseen CARLA frames, SLLCP flags 89.6% of collision-causing trajectories with Qwen3-VL-8B and 88.4% with Cosmos-Reason2-8B, versus 4.6% and 39.1% for the raw binary VLMs. It matters because it offers a planner-agnostic way to reduce missed unsafe trajectories before execution, albeit with more false alarms.