Restoring Linguistic Grounding in VLA Models via Train-Free Attention Recalibration

Researchers have identified a critical failure mode in Vision-Language-Action (VLA) robotic models called 'linguistic blindness,' where robots prioritize visual cues over language instructions when they contradict. They developed ICBench benchmark and proposed IGAR, a train-free solution that recalibrates attention to restore language instruction influence without requiring model retraining.

• →VLA robotic models suffer from 'linguistic blindness,' executing visually plausible actions even when language instructions contradict the visual scene.
• →ICBench diagnostic benchmark was created to systematically test language-action coupling in robotic models using controlled contradictory instructions.
• →Three major VLA architectures (Pi0, Pi0.5, OpenVLA OFT) showed strong visual bias, frequently succeeding at tasks despite impossible instructions.
• →IGAR (Instruction-Guided Attention Recalibration) provides a train-free solution that rebalances attention without architectural modifications.
• →The approach was validated on 30 LIBERO tasks and real Franka robotic arm, effectively preventing erroneous execution while maintaining performance.