Jun Young Lee

High probe accuracy in VLAs is often an illusion: internal state decoders frequently track ambient scene context rather than the target object itself, hiding causal blind spots that trigger policy failures.