Navigating Latent Worlds: A Hypertextual Taxonomy of World Models as Nonlinear Knowledge Structures

Abstract

World models, learned internal representations that simulate environment dynamics, predict future states, and enable counterfactual reasoning, have emerged as a central construct in the pursuit of autonomous intelligence. Yet existing surveys catalogue world models along purely architectural or application-domain axes, neglecting the epistemological question of how these models organise, traverse, and present knowledge. We argue that hypertext theory provides a uniquely powerful lens for this analysis. This survey introduces the Hypertextual World Model (HWM) framework, a formal taxonomy that reconceptualises world models as nonlinear, multi-resolution knowledge structures characterised by typed links between latent states, branching traversal policies, and reader-agent co-construction of meaning. We formalise three constitutive axes, Linking Topology, Traversal Modality, and Authorial Agency, and systematically classify 43 representative systems published between 2018 and 2026 onto this taxonomy (Table 2). Our analysis reveals four paradigm clusters and identifies critical open problems including hypertextual friction in latent traversal, the coherence trap in long-horizon rollouts, and the absence of provenance-preserving link semantics. We conclude with a research agenda situating world models within the broader hypertext programme of augmenting human interpretive agency over computationally mediated knowledge.