FabriMAE I Trust Myself? Self-Evaluating VLA Action Generation with Markov Attention Entropy

2026-08-17

AI Summary

This study addresses the challenge of self-assessing action reliability in Vision-Language-Action (VLA) models lacking external supervision. We propose the Markov Attention Entropy (MAE) framework, which converts internal attention entropy into reliability scores and integrates a multi-sampling strategy to enable unsupervised self-evaluation and test-time action selection. Furthermore, we introduce the LIBERO-Reflect benchmark to elucidate cross-architectural abstractions in action generation. Experimental results demonstrate that MAE outperforms state-of-the-art baselines across multiple metrics, significantly enhancing the robustness of PI-series models with minimal runtime overhead. Consequently, this work provides an efficient solution for the trustworthy deployment of VLA systems by facilitating reliable autonomous decision-making without requiring additional labeled data or external feedback mechanisms during inference.

Abstract

Vision-Language-Action models (VLAs) integrate visual perception, language instruction, and action generation into end-to-end policies across heterogeneous architectures. However, enabling VLAs to self-evaluate their action generation reliability without external supervision remains a major challenge. Existing methods either rely on expert annotations or estimate uncertainty only from output statistics, largely ignoring internal signals. In this work, we observe that internal visual modality entropy exhibits consistent distinctions between successful and failed tasks across heterogeneous VLAs. Although VLAs' architectures differ in their action generation, we show that they share a common latent action generation abstraction evolving under visual perception, language instruction, and state input, which we formulate as a Conditional Generative Markov Chain. Based on this formulation, we propose MAE (Markov Attention Entropy), a self-evaluation framework that directly converts internal attention signals into architecture-aware reliability scores, and introduce LIBERO-Reflect, a 4,000-episode benchmark combining 2,000 standard episodes and 2,000 challenging episodes across four subsets. Extensive experiments across heterogeneous VLA architectures and diverse scenarios show that MAE consistently outperforms state-of-the-art baselines on AUPR, AUROC, and FPR@95. We further instantiate FabriMAE for verifier-free test-time action selection, showing that MAE-guided multiple sampling improves PI-family robustness on LIBERO-Plus with small observed runtime overhead.

Problem

• Vision-Language-Action models
• Self-evaluation
• Action generation reliability
• Uncertainty estimation

Innovation

• Markov Attention Entropy
• Self-Evaluating VLA
• Conditional Generative Markov Chain
• Verifier-Free Action Selection
• LIBERO-Reflect