Abstract
Vision-Language-Action (VLA) models aim to map multimodal inputs to robot actions. However, existing approaches struggle with complex dynamic scenarios due to uniformly treating all visual tokens and relying on human-selected factors, lacking mechanisms to emphasize task-critical evidence and ignoring underlying factors.
Proposed Method
To address this, we propose LEEVLA, a VLA architecture that explicitly guides the model toward informative regions while preserving the structured evolution of latent world representations.
We introduce Drift-Guided Dynamic Prioritization (DGDP), combining Dynamic Position Prioritization (DPP) with Semantic Drift Guidance (SDG) to direct the VLA agent's attention during training.
Additionally, we present Structured Feature Flow Generation (SFFG), which models how prioritized features evolve in latent space using Prototype-to-Periphery (P2P) prediction and a Mutual-Neighborhood Contrastive (MC) loss to maintain topological consistency among neighborhoods.
Together, DGDP and SFFG form a task-aware "where-how" training framework.
Experiments
Extensive experiments on VLA benchmarks show that LEEVLA consistently outperforms prior methods, confirming that explicit task-evidence guidance and structured latent reasoning are crucial for scalable VLA.
Our code is available at GitHub.
Blogger's Review: LEEVLA significantly enhances VLA model performance in dynamic environments through innovative DGDP and SFFG methods, highlighting the importance of explicit task orientation and structured reasoning in complex scenarios. This research offers new insights into robotic intelligence and deserves attention.