Does VLA Even Know the Basics? Measuring Commonsense and World Knowledge Retention in Vision-Language-Action Models

The paper introduces Act2Answer, a protocol that converts VLM knowledge questions into short tabletop episodes where an agent selects an answer through a single object-placement action. This reduces confounding from low-level control when testing commonsense and factual knowledge in VLAs. In a study of 7 VLA models and 9 VLM baselines, VLAs retained strong performance on simple concepts but showed larger gaps from their source VLMs on richer semantic categories. VQA co-training was associated with better knowledge retention, while answer-relevant signals peaked in middle layers and weakened in upper VLA layers.