Is VLA Reasoning Faithful? Probing Safety of Chain-of-Causation in Autonomous Driving Models
A new study published on arXiv investigates the safety and faithfulness of Vision-Language-Action (VLA) models used in autonomous driving. Researchers found that these models exhibit significant unfaithfulness in their reasoning, with less than half of their generated explanations accurately reflecting the visual scene. The study also highlighted critical safety concerns, such as missed pedestrians and fragile trajectory predictions that are highly sensitive to minor visual changes. AI
IMPACT Highlights critical safety flaws in autonomous driving AI, necessitating improved reasoning faithfulness and robustness.