The Geometry of Empowerment Reframes Agent Control

The paper’s real contribution is its warning label: empowerment is a powerful proxy for optionality, not a universal proxy for usefulness.

• –Potential empowerment identifies states with broad reachable futures, while effective empowerment measures how actively learned skills influence those futures.
• –The theory links empowerment to central states and bottlenecks, offering a principled lens for exploration and skill discovery.
• –In tabular settings, empowerment can lower-bound adaptation to unknown rewards under an uninformed prior.
• –That guarantee weakens sharply for continuous environments and anisotropic reward distributions, where controllable dimensions may not be reward-relevant.
• –The accompanying site provides research details and code, but the work remains primarily a theoretical foundation rather than a production-ready agent framework. [Project site](https://empowerment-geometry.github.io/)