In vision-language-action (VLA) models, action supervision is often viewed as a downstream objective for learning action prediction. This paper instead explores action supervision as a force that shapes inherited multimodal representations. It shows that this shaping has a dual effect: it is essential for forming action-compatible representations, but when action supervision is applied too directly to the inherited multimodal pathway, it can destabilize representations that support language-side processing and object grounding. To address this tension, we introduce Action QFormer, a query-based action-facing interface that utilizes instruction-conditioned queries to reorganize inherited multimodal information into action-facing representations before downstream action generation. In zero-shot sim-to-real navigation, Action QFormer improves average closed-loop task success from 18.8% to 56.3%, raises fixed-instruction action-generation correctness from 22.5% to 75.5%, and nearly eliminates out-of-distribution instruction generations. Further analyses indicate that Action QFormer alters how action supervision shapes inherited multimodal representations, reducing broad upstream rewriting while preserving targeted and sometimes constructive action-supervised adaptation. These results suggest that enhancing VLA performance requires not only stronger pretrained backbones but also better methods for selecting and organizing inherited multimodal information while controlling how it is shaped under action supervision.
Blogger's Review: The introduction of Action QFormer offers a fresh perspective and methodology for improving VLA model performance. Its approach to reorganizing multimodal information and its adaptability under action supervision highlight significant potential for future model designs, especially in scenarios requiring a balance of different task objectives. The success rate improvements in practical applications also provide crucial experimental evidence for related fields.