The paper introduces Representational Empowerment (RepEmp), a criterion for deciding what to represent when modeling the world. Instead of just estimating parameters, agents must choose which representational elements to build, retain, and reuse. RepEmp scores each candidate element by how much it expands the agent's future capacity to model and plan, complementing classic empowerment but applied to internal representations. The authors test this in three experiments: human causal learning, agent-based simulations, and an LLM-augmented planner. In the human experiment, RepEmp predicted participants' choices of abstraction granularity better than information-gain alternatives. Simulations showed RepEmp-guided construction contributed more to structure recovery and cross-task transfer than exploration alone. In the open-vocabulary planning domain, RepEmp built more compact symbolic libraries that also generalized better than baselines. Ablating RepEmp removed these benefits.
Agents benefit from representing what expands future planning capacity
A proposed objective called Representational Empowerment predicts human abstraction choices and improves reusable model construction in simulations and LLM-assisted planning.
Top University
Fei Dai · Hanqi Zhou · Alison Gopnik · Charley Wu
University of California, Berkeley · University of Tübingen · TU Darmstadt
Research Digest··2 min read
In one line
Representational Empowerment, which scores how much candidate representations expand future modeling capacity, predicts what humans and agents choose to represent and builds better libraries.
Read it as
Dai et al. study how bounded agents decide which concepts and structures are worth building into persistent models.
The authors formulate continual model construction as a two-level Curator-Actor architecture.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.