New AI Principle Guides Model Construction for Future Capacity

This research introduces Representational Empowerment (RepEmp), a novel principle designed to guide AI agents in the continuous construction of internal models. Unlike traditional approaches that focus solely on accurately reflecting the current environment, RepEmp prioritizes selecting and curating representational elements that maximize an agent's future capacity for modeling and planning. The core idea is to empower the agent to control its internal representations, enabling more efficient learning and adaptation over time. The framework was implemented using a hierarchical Curator-Actor architecture and tested across various experiments. In causal-learning tasks, human participants and simulations guided by RepEmp showed a preference for abstraction granularities that enhanced goal reachability, outperforming information-gain alternatives. Furthermore, RepEmp-guided construction proved more effective for sufficient structure recovery and cross-task transfer. In open-vocabulary planning, an LLM-augmented Curator built more compact and generalizable symbolic libraries, demonstrating the principle's broad applicability and benefits.