Title: Completion at the Boundary (CaB): Deployable Switching with Completion-Aware Control under Limited Calibration

Abstract: While Vision-Language-Action (VLA) agents possess the capability to follow natural-language instructions, deployed systems currently struggle with a critical operational gap: determining precisely when an instruction has been fulfilled. This challenge is particularly pronounced in short composite commands (e.g., "do A, then B"), where poorly timed transitions can trigger cascading failures in subsequent tasks. Because switching acts as an intervention that alters the instructional context—and thereby influences future actions and observations—completion management is inherently closed-loop.

This study investigates completion protocols within a deployable, low-calibration framework. Motivated by the vastness of open-ended instruction spaces, we enforce strict constraints: no relearning during testing and the use of a single, globally calibrated switching rule derived from the development set and applied unchanged to the test set. Under these limitations, reducing asymmetric boundary evidence to a single scalar value proves fragile, especially when task polarities shift.

To address this, we introduce Completion at the Boundary (CaB). Rather than collapsing evidence, CaB predicts an event-local completion object structured as Boundary-Phase Tokens (Before, Hit, or After), thereby preserving two-sided boundary evidence. The system utilizes two components: CaB-When translates this completion object into a minimal, auditable decision regarding when to switch, while CaB-How leverages the same object to condition action generation, ensuring boundary-stable control during handoffs. Evaluated using an intervention-aware E1/E2 protocol, our results demonstrate that CaB enhances both composite execution and handoff quality on a first-person Minecraft VLA benchmark, achieving these improvements while maintaining matched capacity and deployability constraints.

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