Vision-language-action (VLA) policies increasingly reason before acting, but existing models typically compress that intermediate reasoning into a single representation. When several plans are valid, this commitment hides alternatives that could otherwise be selected or revised. A new arXiv preprint introduces the Generative Plan Latent (GPL), which represents the reasoning state as a low-dimensional conditional distribution over plans—such as which object, grasp, or route to use.
Because the reasoning state is addressable, GPL supports three operations without retraining the VLA: sampling plans for coverage, selecting a plan on request (steering), and reweighting plans based on deployment outcomes. The authors tested all three on SimplerEnv manipulation with the π0 model, and extended the analysis to RoboCasa, LIBERO, and a maze domain.
Steering the plan raised the fraction of executions that follow the request from 0.48 to 0.94, while the same guidance applied to the native action distribution showed no reliable gain. The paper also reports that an explicit plan is most useful when it carries decision information not already determined by the observation—a useful boundary condition for when this kind of reasoning representation matters.