Abstract
Vision–language–action (VLA) models provide a promising interface between semantic reasoning and executable behavior for end-to-end autonomous driving, yet sparse closed-loop feedback leaves the decisive time step and responsible component of a structured driving decision unspecified. This ambiguity weakens credit assignment and can lead to broad policy updates that disturb previously reliable behaviors. We propose FAR-MindDrive, a failure-aware risk attribution framework that converts closed-loop failures into localized corrective supervision before policy repair. FAR-MindDrive first performs event-centered temporal backtracking to identify the short decision history most relevant to a violation, then applies factorized counterfactual interventions to longitudinal and lateral meta-actions and measures their event-specific geometric risk reduction. High-confidence corrections update the Decision Expert, while imitation replay and reference-policy regularization constrain unnecessary drift. On Bench2Drive, FAR-MindDrive achieves a five-run mean Driving Score of 79.12 and Success Rate of 57.73%, compared with 75.85/49.30% for MindDrive-IL and 78.04/55.09% for MindDrive; the mean collision rate is 5.91%. These results show that structured failure attribution enables targeted, interpretable, and resource-efficient VLA post-training while retaining the deployed action-generation architecture.