Perturbation-Based Uncertainty for Failure Detection in Vision-Language-Action Models

Vision-Language-Action (VLA) models have shown strong performance in robotic manipulation, but reliable uncertainty quantification remains challenging, particularly under distribution shift. Unlike autoregressive policies, many modern VLA models generate continuous actions through regression or flow-based generation, where explicit predictive probabilities are unavailable. Moreover, existing approaches often rely on stochastic action sampling or supervised failure labels, limiting their applicability across diverse pretrained VLA models. In this work, we propose a label-free and model-agnostic framework for inference-time uncertainty estimation through hidden activation perturbations, motivated by Bayesian perspectives on local model variations. Specifically, we inject Gaussian perturbations into transformer hidden activations and estimate epistemic signals from disagreement across perturbed action predictions. Experiments on LIBERO and LIBERO-PRO show that perturbation-based uncertainty consistently improves failure detection under distribution shift compared to sampling-based uncertainty, providing a practical uncertainty signal for VLA models.

Why Linear Interpretability Works: Invariant Subspaces as a Result of Architectural Constraints

Linear probes and sparse autoencoders consistently recover meaningful structure from transformer representations -- yet why should such simple methods succeed in deep, nonlinear systems? We show this is not merely an empirical regularity but a consequence of architectural necessity: transformers communicate information through linear interfaces (attention OV circuits, unembedding matrices), and any semantic feature decoded through such an interface must occupy a context-invariant linear subspace. We formalize this as the \emph{Invariant Subspace Necessity} theorem and derive the \emph{Self-Reference Property}: tokens directly provide the geometric direction for their associated features, enabling zero-shot identification of semantic structure without labeled data or learned probes. Empirical validation in eight classification tasks and four model families confirms the alignment between class tokens and semantically related instances. Our framework provides \textbf{a principled architectural explanation} for why linear interpretability methods work, unifying linear probes and sparse autoencoders.