Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This study investigates the relationship between corruption-induced embedding drift and task-aligned hazard scores in vision-language models (VLMs) for autonomous driving hazard detection. By analyzing controlled corruptions on BDD100K road scenes, the authors reveal that the coupling between representation drift and decision drift is highly dependent on the type of corruption, with some leading to significant decision instability despite minor embedding changes. The findings highlight the need for robustness benchmarks to incorporate task-aligned stability measures, as different corruption types can lead to asymmetric failure modes in hazard detection.
Corruption types can drastically alter hazard detection stability, revealing that minor embedding changes can lead to major decision failures in autonomous driving systems.
Vision-language models (VLMs) are increasingly used for scene understanding in autonomous driving, but robustness analysis often relies on task-agnostic embedding stability alone. We study whether corruption-induced embedding drift predicts changes in a task-aligned hazard score derived from CLIP image-text similarities. Using controlled corruptions on BDD100K road scenes, we compare embedding drift against margin drift, defined as the change in hazard score under perturbation. The relationship is highly corruption-dependent: some families exhibit strong coupling between representation drift and decision drift, while others induce hazardous decision instability despite relatively modest embedding change. Furthermore, corruption families differ in failure direction: most suppress hazard detections via false negatives, while occlusion instead triggers false alarms, suggesting that benchmark design should account for asymmetric failure modes, not just overall instability rates. These results suggest that robustness benchmarks should include task-aligned stability measures in addition to embedding-level perturbation statistics.