LEEVLA Framework Guides Robot Vision-Language-Action Models to Focus on Task-Critical Visual Evidence

Researchers have introduced LEEVLA, a vision-language-action (VLA) architecture that explicitly guides robot models to attend to task-critical visual regions while preserving structured latent-world representations, improving performance on robotic manipulation benchmarks. VLA models map multimodal inputs to robot actions but typically treat all visual tokens uniformly and rely on human-selected reasoning factors, limiting their ability to identify task-relevant evidence in dynamic scenes. LEEVLA addresses this with two mechanisms: drift-guided dynamic prioritization (DGDP), which combines dynamic position prioritization with semantic drift guidance to direct attention to informative regions during training, and structured feature flow generation (SFFG), which models how prioritized features should evolve in latent space via prototype-to-periphery prediction and a mutual-neighborhood contrastive loss to maintain topological consistency. Together, DGDP and SFFG form a "where-how" training framework that tells the model where to attend and how those attended features should evolve. Experiments on VLA benchmarks show LEEVLA consistently outperforms prior methods, confirming that explicit task-evidence guidance and structured latent reasoning are both crucial for scalable vision-language-action models. The researchers have released their code publicly.