Do Vision-Language-Action Models Understand Instructions? A Mechanistic Interpretability Study on Language Grounding

This paper presents a mechanistic interpretability study on the language grounding capabilities of two state-of-the-art Vision-Language-Action models, π₀.₅ and GR00T N1.7. The study applies activation and attribution patching to analyze how these models respond to variations in task instructions, including synonym replacement, semantic scaling, directional corruption, random object substitution, and empty strings. Results show differing sensitivities to abstract rephrasing and directional language, with distinct patterns of activation in each model.