Determine whether the vision pathway exhibits analogous causal and representational dissociations

Determine whether the vision pathway in the [?]$_{0.5}$ and GR00T N1.7 Vision-Language-Action models exhibits causal and representational dissociations analogous to those identified for language perturbations.

Background

The paper combines causal patching methods with CKA to show that causal effects and representational changes can dissociate, most clearly for directional language in GR00T N1.7. The analysis is restricted to perturbations of language instructions. The authors leave unresolved whether comparable dissociations occur in the visual processing pathway and propose applying the same perturbation framework to vision as a test.

References

Additionally, applying our perturbation framework to vision would test whether the vision pathway shows analogous causal/representational dissociations.

— Do Vision-Language-Action Models Understand Instructions? A Mechanistic Interpretability Study on Language Grounding  (2610.10178 - Wulff et al., 7 Oct 2026) in Section 6, Future Work