Title: Looking Past Task Completion: Behavioral and Representational Analysis of WAM and VLA Models

Abstract:

Vision-language-action (VLA) policies and World-Action Models (WAMs) have emerged as two pivotal frameworks in the field of robotic manipulation. Despite their growing prominence, it remains an open question whether the future prediction capabilities inherent in WAMs yield behavioral enhancements that extend beyond mere task completion. This study investigates whether WAMs simply incorporate future forecasting or if they fundamentally alter both robot behavior and internal neural representations to provide actionable insights for control systems. To address this, we propose a model-agnostic diagnostic framework that evaluates WAMs and VLAs through two distinct but complementary methodologies: behavioral rollout analysis and feature analysis utilizing sparse autoencoders.

The behavioral assessment protocol focuses on four key metrics: consistency in action dynamics, progress toward target objects, resilience against distractor disturbances, and computational runtime costs. Concurrently, the feature-space protocol categorizes internal representations as either memorized, reactive, or predictive, thereby determining if models successfully encode future-oriented structural information. We benchmark seven distinct policies across the LIBERO and RoboTwin2.0 datasets, encompassing direct VLAs as well as joint, sequential, and auxiliary WAM configurations.

Our findings indicate that relying solely on success rates obscures significant underlying differences. While WAMs frequently demonstrate improvements in object-level behavior and target selectivity, these advantages are contingent upon the specific architectural design and come with increased inference overhead. Notably, sequential WAMs exhibit the most distinct predictive structures, whereas auxiliary and joint WAMs tend to either compress or entangle future information, respectively. These insights point toward new avenues for designing WAM architectures that maintain behaviorally actionable future representations, ultimately facilitating more efficient robotic manipulation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC