VLA-Trace: Diagnosing Vision-Language-Action Models through Representation and Behavior Tracing

2026-05-28

AI Summary

This work systematically investigates how vision-language-action (VLA) models translate multimodal knowledge into embodied control mechanisms. To this end, we introduce an integrated diagnostic framework that combines centered kernel alignment (CKA) for cross-modal representation alignment, attention knockout interventions, and behavior-level probing. Applying this framework, we uncover critical differences among prominent VLA models—such as π₀.₅ and OpenVLA—pertaining to fine-tuning adaptation, multimodal routing, and semantic instruction following. Our analysis reveals that while current models effectively generate visual trajectories, they exhibit notable limitations in adhering to fine-grained language instructions. These findings provide empirical grounding and concrete directions for the future design of VLA architectures.

Abstract

Understanding how Vision-Language-Action (VLA) models transform multimodal knowledge into embodied control remains an open challenge. We present VLA-Trace, a progressive diagnostic framework that analyzes VLA models through a unified evidence chain from representation dynamics to causal control attribution and behavioral manifestation. It specifically combines cross-modal and checkpoint-drift centered kernel alignment (CKA) to trace representation evolution, attention knockout interventions to identify modality-specific control pathways, and rollout-level behavioral probes to examine grounding, shortcut dependence, and semantic following. Experiments on $π_{0.5}$ and OpenVLA reveal three key findings. First, the two models exhibit distinct modality-specific adaptation dynamics during VLA finetuning. Second, they rely on different multimodal routing strategies and layer-wise dependencies during action decoding. Third, although VLA policies excel at visually grounded trajectory generation, they remain limited in fine-grained semantic following. These findings highlight future directions for representation-preserving adaptation, causal VLA circuits, and compositional semantic control.

Problem

• Vision-Language-Action models
• multimodal knowledge
• embodied control
• representation dynamics
• semantic following

Innovation

• VLA-Trace
• representation tracing
• attention knockout
• behavioral probing
• multimodal alignment