Research questionHow can world models represent entities and causal interactions for reasoning and decision-making?Predictive world models may generate plausible outcomes without representing the entities, properties, and interactions that explain those outcomes. This makes it difficult to use them for causal reasoning, planning, and action beyond the training distribution. Latest papersRecent research connected to this question, newest first.A Unifying Perspective on Causal World Models: From Observations to Representations to StructureThe source presents a formal, conceptual treatment of causal world models spanning perceptual observations, representations, and environmental structure. It addresses entity properties, entity-to-entity and entity-to-environment interactions, and identifiability of model components, but does not establish a particular implementation or empirical deployment result.research paper · Sep 3, 2026AGI Maze Prediction Datasets: A Compact Benchmark for Learning World Dynamics with TransformersEvidence comes from procedurally generated, stateful grid worlds involving per-step transition prediction, fixed-horizon state prediction, and sequential textual-observation prediction. The benchmark uses source-maze-disjoint training and validation splits with greedy exact-match evaluation, comparing from-scratch byte-level Transformers with latent-memory and spatial-memory variants. The evidence is limited to these controlled tasks and reported architectures.research paper · Sep 2, 2026What Is Worth Representing? Representational Empowerment for Continual Model ConstructionThe evidence covers closed-vocabulary causal-learning tasks with human participants and matched simulations, plus an open-vocabulary planning domain using an LLM-augmented Curator. The studied system uses a hierarchical Curator–Actor architecture and evaluates persistent representational libraries for structure recovery, transfer, compactness, and generalization.research paper · Sep 2, 2026