Can robots learn to fix bad advice without rewriting the rulebook?

Robots often inherit flawed policies from pretrained models or human operators. GLOVES adapts these actions on-the-fly by learning a flow model that transports non-expert actions toward an expert distribution, then uses reverse-flow scoring to decide which actions to correct and which to trust. The result: improved task success while preserving the agent's original intent, trained on minimal expert data. Code and demos released.