Research map / Multi-agent systems and games / Human-AI and goal-conditioned

What Is Worth Representing? Representational Empowerment for Continual Model Construction

NeurIPS 2026 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: World-model-based agents · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes representational empowerment, an information-theoretic criterion scoring candidate representational elements by how much they expand future modeling and planning capacity, realized in a hierarchical Curator-Actor architecture and tested in human causal learning, simulations, and LLM-driven PDDL library construction.

Problem

Agents must decide which representational elements to build into a task-specific model and retain in a persistent cross-environment library under resource limits; existing continual learning, world-model, and causal learning methods assume a fixed vocabulary and lack a criterion for this choice.

Approach

Treat edits to the agent's own representation as the input channel of an information-theoretic capacity (representational empowerment), grounded by goal-achievement signatures and cost-regularized, so that elements are valued by how much they expand plannable goals per unit cost rather than by uncertainty reduction or external state diversity.

Key results

• In Exp. 3, the Curator reaches backward transfer .61 and held-out .47 on BabyAI versus .40 and .35 for WorldCoder, with 5.14M tokens vs 6.50M and library 8.0 KB vs 18.3 KB.
• In Exp. 2, the factored RepEmp-granularity plus EnvEmp-intervention condition reaches held-out F1 .703, versus .383 for InfoGain and .365 for Random at the granularity level.
• In Exp. 1 (n=70), participants chose medium granularity on 43% of decisions, and greater medium use correlated with more goals achieved (r=.24, p=.042).

Surprising findings

• Cost minimization alone would favor coarse granularity, yet participants moved away from coarse toward medium over tasks.
• WorldCoder exceeds the Curator on Zelda forward transfer (.35 vs .24) despite failing on backward and held-out transfer.

Limitations

• Open-vocabulary experiments use only two grid-world types and are described as directional rather than benchmark results.
• The LLM curator can only score candidates its proposer generates; a failure case in BabyAI held-out had all methods below 15% success.
• Human experiment uses a fixed set of interventions and granularities, with only behavioral and model-based evidence.

Experimental setup

Datasets: BabyAI Zelda (EMPA layout) procedurally generated SEM causal-learning curricula

Benchmarks: BabyAI Zelda alchemy causal-discovery environment

Metrics: F1 for graph recovery delta log-likelihood held-out forward/backward/held-out transfer success LLM tokens library size (KB) goal achievement

Baselines: PPO Motif WorldCoder Hoarder InfoGain EnvEmp Random

Builds on: environmental empowerment information gain WorldCoder Motif PDDL

Closest papers by meaning

• Do LLMs Build World Representations? Probing Through the Lens of State Abstraction (NeurIPS 2024 · similarity 0.86)
• Curious Causality-Seeking Agents Learn Meta Causal World (NeurIPS 2025 · similarity 0.86)
• WorldCoder, a Model-Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment (NeurIPS 2024 · similarity 0.85)
• Plasticity as the Mirror of Empowerment (NeurIPS 2025 · similarity 0.85)
• Learning 3D Persistent Embodied World Models (NeurIPS 2025 · similarity 0.85)
• Language Agents Meet Causality -- Bridging LLMs and Causal World Models (ICLR 2025 · similarity 0.85)