VLADriveBench: Evaluating CoT-Action Relationship in VLA for Autonomous Driving

Researchers have developed VLADriveBench, a new evaluation framework that assesses whether the chain-of-thought reasoning produced by vision-language-action (VLA) models for autonomous driving is relevant and causally connected to the driving action, rather than just measuring trajectory quality. Testing on three models across two architectures revealed that observational metrics and causal analysis can produce sharply divergent results, with one model scoring high on alignment yet having epiphenomenal reasoning, while another scored lower yet produced strongly causal chain-of-thought. The findings highlight the need for evaluating the causal relationship between reasoning and action in autonomous driving systems, as visual salience was found to gate the extent of chain-of-thought influence.

arXiv:2606.12706v1 Announce Type: new Abstract: Vision-language-action VLA models generate chain-of-thought CoT reasoning alongside driving trajectories, but existing benchmarks evaluate only trajectory quality and do not assess whether the CoT is relevant, consistent, or causally connected to the driving action. We introduce VLADriveBench, a framework that combines observational metrics mentioning, hallucination, contradiction, action alignment with a CoT intervention protocol to provide complementary views of the CoT-action relationship. Applying VLADriveBench to three models across two architectures, we find that the two analyses can diverge sharply: ORION scores highest on observational alignment yet its CoT is epiphenomenal, while Alpamayo v1.5 scores lower yet its CoT is strongly causal, with visual salience gating the extent of CoT influence.