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.

PaperTop Universitycs.LGarXiv:2609.02322v1 ↗

Fei Dai · Hanqi Zhou · Alison Gopnik · Charley Wu

University of California, Berkeley · University of Tübingen · TU Darmstadt

Research Digest·4 September 2026·2 min read

Read the paper →

What they did

The authors formulate continual model construction as a two-level Curator-Actor architecture. The Actor uses an environment-specific model, while the Curator maintains a persistent library of reusable representational elements; Representational Empowerment (RepEmp) scores elements by how much they expand future modeling and planning capacity.

They evaluated the framework in three settings: a human causal-learning task with fixed representational options, matched simulations separating model construction from exploration, and an open-vocabulary planning domain where an LLM-assisted Curator created symbolic libraries.

Key findings

Human participants selected abstraction levels that increased goal reachability rather than simply maximizing fidelity to the underlying world; RepEmp predicted this pattern better than information-gain alternatives.
In matched simulations, RepEmp-guided representation construction contributed more than exploration to recovering sufficient causal structure and transferring it across tasks.
In open-vocabulary planning, the LLM-assisted Curator produced symbolic libraries that were both more compact and more generalizable than baseline libraries.
Removing RepEmp eliminated the reported compactness and generalization benefits, indicating that they did not arise from the architecture or LLM alone.

Why it matters

The work shifts model learning from only estimating parameters or causal relations to deciding which representational building blocks deserve limited storage and computation. This offers a principled objective for agents that must accumulate reusable concepts across changing environments rather than reconstructing a full model for every task.

Caveats

The abstract does not report sample sizes, effect sizes, baseline details, or the breadth of the planning environments, so the empirical scale and robustness cannot be assessed here. The evidence comes from controlled causal-learning and symbolic-planning settings; whether RepEmp remains effective in noisy, high-dimensional, real-world domains is open.