Addressing Linguistic Blindness in VLA Models via Attention Recalibration

VLA models often prioritize visual priors over language instructions, a failure mode termed linguistic blindness. This research introduces ICBench to evaluate language-action coupling and proposes a train-free attention recalibration method to restore linguistic grounding in robotic policies under out-of-distribution instructions.