Embodied Interpretability: Linking Causal Understanding to Generalization in Vision-Language-Action Models

2026-04-30

AI Summary

Vision-language-action (VLA) policies often exhibit poor generalization under distribution shifts, relying on spurious visual correlations rather than task-relevant causal factors. This work frames visual-action attribution as a causal intervention problem and introduces two novel metrics: Interventional Saliency Score (ISS) and Noise-to-Signal Quality Ratio (NMR). To the best of our knowledge, this is the first application of causal intervention techniques to diagnose causal misalignment in embodied intelligence. By leveraging interventional masking, causal inference, and unbiased estimation, the proposed approach yields more faithful explanations of model decisions. Experiments demonstrate that NMR effectively predicts out-of-distribution generalization performance, while ISS outperforms existing attribution methods, offering a new diagnostic tool for enhancing the interpretability and reliability of VLA policies.

Abstract

Vision-Language-Action (VLA) policies often fail under distribution shift, suggesting that decisions may depend on spurious visual correlations rather than task-relevant causes. We formulate visual-action attribution as an interventional estimation problem. Accordingly, we introduce the Interventional Significance Score (ISS), an interventional masking procedure for estimating the causal influence of visual regions on action predictions, and the Nuisance Mass Ratio (NMR), a scalar measure of attribution to task-irrelevant features. We analyze the statistical properties of ISS and show that it admits unbiased estimation, and we characterize conditions under which action prediction error provides a valid proxy for causal influence. Experiments across diverse manipulation tasks indicate that NMR predicts generalization behavior and that ISS yields more faithful explanations than existing interpretability methods. These results suggest that interventional attribution provides a simple diagnostic approach for identifying causal misalignment in embodied policies.

Problem

• Vision-Language-Action
• distribution shift
• causal understanding
• spurious correlations
• generalization

Innovation

• Interventional Attribution
• Causal Interpretability
• Vision-Language-Action Models
• Generalization Diagnosis
• Nuisance Mass Ratio