LEEVLA: Seeing What Matters in Latent Environment Evolution for Vision-Language-Action

Jul 16, 2026

Vision-language-action (VLA) models map multimodal inputs to robot actions, yet many methods fail in complex dynamic scenes because they treat visual tokens uniformly and rely on human-chosen factors, offering limited emphasis on task-critical evidence and overlooking latent underlying factors. LEEVLA introduces drift-guided dynamic prioritization to steer attention to salient, instruction-relevant regions and structured feature flow generation to model their evolution in latent space. Experiments on VLA benchmarks show consistent performance gains, highlighting scalable task-evidence guidance and latent structured reasoning.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

Abstract

Introduction

Background and motivation

Limitations of prior context-clue reasoning

LEEVLA approach: latent reasoning and task evidence

LEEVLA: Seeing What Matters in Latent Environment Evolution for Vision-Language-Action Qi Lyu 1 ,2 ,3 , Baicheng Liu 1 ,2 , Xudong Wang 1 ,2 ,3 , Jiahua Dong4 , Lianqing Liu 1 ,2 , Zhi Han 1 ,2 1 State Key Laboratory of Robotics and Intelligent Systems 2 Shenyang Institute of Automation, Chinese Academy of Sciences 3University of Chinese Academy of Sciences 4Mohamed bin Zayed University of Artificial Intelligence arXiv :2607 .08 182v 1 [ cs .CV] 9 Jul 2026 Abstract Vision-language-action (VLA) models aim to map multimodal inputs to robot actions. However, most existing approaches struggle to cover complex dynamic scenarios due to treating all visual tokens uniformly and reasoning with human-selected factors, which lack mechanisms to emphasize task-critical evidence and ignore underlying factors. To address this issue, we propose LEEVLA, a VLA architecture for seeing what matters in Latent Environment Evolution that explicitly guides the model toward informative regions while preserving the structured evolution of latent world representations. To identify salient and instruction-relevant regions, we introduce drift-guided dynamic prioritization (DGDP), which combines dynamic position prioritization (DPP) with semantic drift guidance (SDG) to guide the VLA agent where to attend during training. On top of this, we introduce structured feature flow generation (SFFG), which models how these prioritized features should evolve in latent space via prototypeto-periphery (P2P) prediction, and a mutual-neighborhood contrastive (MC) loss to maintain topological consistency among neighborhoods. Together, DGDP and SFFG forma task-aware “where–how” training framework. Extensive experiments on VLA benchmarks show that LEEVLA consistently outperforms prior methods, confirming that explicit task-evidence guidance and structured latent reasoning are both crucial for scalable VLA. Our code is available in the [https://github.com/LyuQi127/LEEVLA](https://github.com/LyuQi127/LEEVLA). 1. Introduction Vision-Language-Action (VLA) [3, 8, 11, 15, 18, 23, 29, 42, 53, 58, 63] models aim to ground perception and language in action, mapping observation and language instructions directly to low-level controls in the closed loop. | | Multimodal inputs Instruction: pick up the black bowl | | | | | Actions | | | | 􀜼௜ Known factors 􀜼௝ Unknown factors | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | | | 􀜼௜ (a) Causal graph | | | 􀜽௧ାଵ | | | | | | | | | | | | | | | | | | | | | | | | | | 􀜽௧ା் | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Sample | | | | | Sample | | | | | | | | Context Clues Action | | | | | Context Clues Action | | | | | | | | | | | | | | | | | | | | | | | | | | LEEVLA | | | | | | | | | | | | | | | | | | | | | Multimodal Action inputs tokens (b) Prior methods | | | | | Multimodal Action inputs tokens (c) Our method | | | | | | | Figure 1 . Comparison between our method and prior methods.(a): Causal graph between multimodal inputs and action chunk.(b): Prior methods: VLA reasons on factors of specific context clues. (c): Our method: VLA reasons on abstract context clues. By jointly encoding visual observations, proprioception, and language instruction, VLAs offer a path toward robots that can perceive, reason, and act in open environments. Recent research typically employs large language models (LLMs) or visual-language models (VLMs) to construct VLAs [31, 35, 43, 46], and then trains these VLAs on largescale robot demonstrations or simulated interaction traces [28, 34] . Notably, the generation of action is governed by both known and unknown factors [47], as illustrated by the causal graph in Fig. 1 (a) . Recently,

FAQs

What problem do existing VLA approaches struggle with in dynamic scenarios?

They often reason with visual tokens treated uniformly and depend on human-selected factors, which weakens the focus on task-critical evidence and ignores underlying latent factors.

How does LEEVLA decide where the VLA agent should attend during training?

It uses drift-guided dynamic prioritization (DGDP), combining dynamic position prioritization with semantic drift guidance to highlight salient, instruction-relevant regions.

What training mechanisms does LEEVLA use to model how prioritized features evolve?

It introduces structured feature flow generation (SFFG) using prototype-to-periphery prediction in latent space and a mutual-neighborhood contrastive loss to preserve neighborhood topological consistency.

Jul 16, 2026

Upload by Theodora