Gopnik coauthors 'Representational Empowerment' for AI agents

Humans building causal models don't try to mirror the world; they build models tuned to whichever goal they can reach. That is the empirical claim at the center of a new preprint from Fei Dai, Hanqi Zhou, Alison Gopnik and Charley Wu, posted to arXiv on 2 September 2026.

The authors introduce a scoring rule they call Representational Empowerment, or RepEmp. Instead of asking how much a candidate element reduces uncertainty about the environment, RepEmp scores elements by "how much they expand the agent's future capacity to model and plan," redefining the classic empowerment idea as "control over internal representations instead of external states." The framework is implemented as a hierarchical Curator-Actor architecture.

Three experiments follow. In a closed-vocabulary causal-learning task, human participants "construct causal models at varying abstraction granularities to maximize goal reachability rather than fidelity to the world," a signature the paper reports is "better predicted by RepEmp than by information-gain alternatives." Matched simulations found that RepEmp-guided construction "contributes more than exploration to sufficient structure recovery and cross-task transfer." In an open-vocabulary planning domain, an LLM-augmented Curator "builds more compact symbolic libraries, which also generalize better than baselines." Ablating RepEmp erased those gains.

Two researchers we track shared the link in the first days after it went up.