Vision-language-action (VLA) models are designed to follow natural-language instructions in robotic or embodied settings, inheriting that ability from their language backbones. But most evaluations have concentrated on what-level instructions—those that specify the goal or task—rather than how-level instructions that describe the way a task should be performed.

This paper from arXiv directly addresses that gap. It asks whether VLAs really follow how-level instructions and presents a diagnosis of their performance on this dimension. The abstract notes that prior work has largely ignored this level of instruction, suggesting that current benchmarks may not capture how precisely these models adhere to user intent.

The study is positioned as a diagnosis rather than a new method, which could help the field identify where VLA behavior diverges from language-model-like instruction following. As with any single preprint, the findings are not yet peer-reviewed, but the question it raises is central to deploying VLAs in settings where how a task is done matters as much as what is done.