Event
AWARE: A Training-Free, Plug-and-Play Observation Evidence Retrieval Module for VLA
1 reports · 1 sources
Jiabing Yang and other authors propose AWARE, a training-free, plug-and-play module for vision-language-action models (VLA). It tracks layer-wise uncertainty via element-wise action entropy and, using associative memory lookup, retrieves observation evidence into the key-value memory of the feedforward network. The authors claim that, without changing the architecture or additional training, AWARE consistently improves the performance of multiple VLA architectures at very low computational cost, with the largest gains on long-horizon tasks. These results are currently the authors' claims in the paper.
Report timeline
- AWARE: Action-Entropy-Weighted and Attention-Guided Retrieval of Observation Evidence for Vision-Language-Action Models
The authors propose AWARE, a training-free, plug-and-play module for Vision-Language-Action (VLA) models that tracks layer-wise uncertainty via element-wise action entropy and retrieves observation evidence into the feed-forward networks' key-value memory through associative memory lookup.
arXiv cs.RO ·