Vision-language models are increasingly used for driving-scene understanding, but most evaluations assume well-organized traffic environments. A new preprint argues that defensive reasoning under weakly structured Global South traffic remains poorly understood, leaving a critical gap in real-world safety assessment.

To address this, the authors introduce BD-HazardVLM, a diagnostic benchmark built from 500 road scenes in Bangladesh. The benchmark is designed to test whether models can identify latent hazards—threats that are present but not immediately obvious—rather than simply recognizing visible objects or events.

The work highlights a blind spot in current model evaluation. Without benchmarks like this, it is difficult to know whether vision-language models can support safe driving decisions in settings where lane markings, signals, and predictable vehicle behavior are often absent.