Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models

The source text describes a fine-tuned Vision-Language-Action (VLA) policy that reliably places an alphabet soup in a basket as seen during fine-tuning, but fails when the basket is shifted about five centimeters, dropping the soup off the table. The authors present an interpretability pipeline aimed at understanding what such a policy computes internally. The excerpt is truncated before its methods and findings are stated.

Key takeaways

• A fine-tuned VLA policy performs a pick-and-place task reliably under conditions seen during fine-tuning.
• A small shift of the basket, about five centimeters, causes the policy to drop the soup off the table.
• The authors developed an interpretability pipeline to investigate what the policy is computing internally.

Why it matters

The failure under a small spatial shift suggests fine-tuned robot policies may not generalize as expected, and interpretability tools could help explain why.

Read with care

The supplied text is truncated, so the pipeline's methods and results are not visible; source claims not independently checked.