ATA: Bridging Implicit Reasoning with Attention-Guided and Action-Guided Inference for Vision-Language Action Models
Cheng Yang, Jianhao Jiao, Lingyi Huang, Jinqi Xiao, Zhexiang Tang, Yu Gong, Yibiao Ying, Yang Sui, Jintian Lin, Wen Huang, Bo Yuan
AI summary
Problem
Existing reasoning-enhanced VLA models rely on costly, data-intensive annotations and require retraining, leading to longer inference times and reduced efficiency, while pure VLA models lack robustness in complex tasks.
Approach
ATA injects implicit reasoning at inference time by adaptively refining visual inputs using two complementary strategies: an attention-guided mask that highlights task-relevant regions from the model's internal attention maps, and an action-guided mask that emphasizes the robot's intended motion direction.
Key results
- Training-free, plug-and-play framework requiring no extra annotations or retraining
- Consistently improves task success and robustness across multiple VLA models
- Enhances inference efficiency by reducing effective reasoning horizons and preventing cascading errors
- Achieves up to 10% performance improvement in real-world block-stacking and significant gains in simulation benchmarks
Why it matters
Enables scalable, efficient, and robust deployment of large VLA models in real-world robotics without the computational and data overhead of training-based reasoning methods.
Abstract
Vision-Language-Action (VLA) models rely on current observations, including images, language instructions, and robot states, to predict actions and complete tasks. While accurate visual perception is crucial for precise action pre- diction and execution, recent work has attempted to further improve performance by introducing explicit reasoning during inference. However, such approaches face significant limitations. They often depend on data-intensive resources such as Chain- of-Thought (CoT) style annotations to decompose tasks into step-by-step reasoning, and in many cases require additional visual grounding annotations (e.g., bounding boxes or masks) to highlight relevant image regions. Moreover, they involve time- consuming dataset construction, labeling, and retraining, which ultimately results in longer inference sequences and reduced efficiency. To address these challenges, we propose ATA, a novel training-free framework that introduces implicit reasoning into VLA inference through complementary attention-guided and action-guided strategies. Unlike CoT or explicit visual- grounding methods, ATA formulates reasoning implicitly by integrating attention maps with an action-based region of interest (RoI), thereby adaptively refining visual inputs without requiring extra training or annotations. ATA is a plug-and- play implicit reasoning approach for VLA models, lightweight yet effective. Extensive experiments show that it consistently improves task success and robustness while preserving, and even enhancing, inference efficiency.