Title: Continuous Reasoning for Vision-Language-Action

Abstract:

While natural language serves as a potent reasoning tool for language and vision-language models, it is ill-suited for the fine-grained demands of continuous control. Textual inputs and explicit subgoals function at a task-level resolution, yet Vision-Language-Action (VLA) policies require decisions at a significantly finer temporal scale. Consequently, a single reasoning step may encompass numerous action chunks while maintaining only a loose connection to the immediate action required. This discrepancy prompts a fundamental question for VLA systems: what medium should effectively replace language?

We posit that an effective reasoning medium for VLA must satisfy three criteria: it must be shareable across different model instances, verifiable via improvements in downstream actions, and aligned with the structure of temporally extended control. Guided by this perspective, we introduce Continuous Reasoning for Vision-Language-Action. Our approach first generates continuous reasoning represented as a structured set of "continuous thoughts," which are subsequently reused as shared context to guide chunk-structured action generation.

However, superior action prediction does not inherently validate the quality of reasoning. If the internal medium cannot be shared across instances or independently verified through enhanced downstream control, the added latent variables may merely serve as model-specific shortcuts that aid in memorizing observed behaviors rather than enabling generalizable control. To address this, we define continuous reasoning as a shared Gaussian latent interface. We train this interface using a self-verification objective, wherein an exponential-moving-average teacher model must successfully interpret the student’s reasoning when predicting target actions.

Our empirical results demonstrate that Continuous Reasoning enhances robustness on LIBERO-PRO and achieves strong performance on physical robots. Specifically, it increased mean subtask success rates by 40.4% on TX-G2 (an AgiBot G2-compatible variant) and by 26.3% on HSR, compared to the $\pi_0.5$ baseline. These findings suggest that effective reasoning in VLA systems depends less on adding extra tokens and more on establishing a shareable, verifiable internal language for action.

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