AIBullisharXiv – CS AI · Mar 36/108
🧠
ATA: Bridging Implicit Reasoning with Attention-Guided and Action-Guided Inference for Vision-Language Action Models
Researchers propose ATA, a training-free framework that improves Vision-Language-Action (VLA) models through implicit reasoning without requiring additional data or annotations. The approach uses attention-guided and action-guided strategies to enhance visual inputs, achieving better task performance while maintaining inference efficiency.