What to Test Next: Interpretable Coverage Gap Discovery in Driving VLMs
Researchers have developed SliceNav, an LLM-orchestrated pipeline designed to improve the testing of driving vision-language models (VLMs). This system addresses the challenge of sparse verification by recommending under-tested regions within Operational Design Domains (ODDs). SliceNav utilizes a scoring rule that prioritizes rare conditions and propagates risk from similar tested scenarios, ensuring deterministic and auditable validation processes. AI
IMPACT This research offers a novel approach to systematically identify and address weaknesses in driving vision-language models, potentially leading to safer autonomous systems.