The notion that intelligent agents maintain internal models of their environment traces to Craik's hypothesis that organisms carry “small-scale models” of external reality [15], formalised by Johnson-Laird's mental model theory [33]. In AI, Minsky's frames [47] and Sutton's Dyna architecture [57] established the model-based RL paradigm.

The modern era began with Ha and Schmidhuber's 2018 “World Models” [24], combining a VAE [35] with an RNN for policy training via latent dreaming. PlaNet [26] extended this to continuous control from pixels. The Dreamer series [25, 27, 28] scaled this via the Recurrent State-Space Model (RSSM), with DreamerV3 demonstrating a single architecture mastering 150+ tasks. SimPLe [34] showed that even simple video-prediction models can serve as effective world models for Atari.

A critical definitional question has since emerged. Gupta [23] identifies three “essential properties” distinguishing genuine world models from pattern completers: persistent state, physical constraint enforcement, and closed-loop utility. Under these criteria, pure video generators like Sora [53], which predict plausible pixel sequences but lack persistent state or action-conditioned branching, may not qualify as world models in the strong sense. This debate maps directly onto our taxonomy: it motivates our distinction between systems that merely narrate futures (low agency, linear topology) and those that equip agents to navigate them.

The field has diversified rapidly. DIAMOND [1] demonstrated that diffusion-based world models preserve visual details critical for RL, achieving state-of-the-art on Atari 100k (mean HNS of 1.46). IRIS [45] and Δ -IRIS [46] showed that tokenized world models, discretising latent space into visual token vocabularies composed by Transformers, can match or exceed continuous-latent approaches. TD-MPC2 [30] demonstrated scalable world models for continuous control. State-space models (SSMs) like Mamba [22] offer linear-time sequence modelling that does not fit cleanly into existing categories. Flow matching [40] has emerged as an alternative to diffusion for trajectory generation.

At the foundation scale, Genie 3 [21] represents the current frontier: a general-purpose interactive world model generating photorealistic environments at 720p and 24 fps from text prompts, with visual memory extending approximately one minute. DeepMind reports that Genie 3 “teaches itself how the world works” through self-supervised learning rather than hard-coded physics, and has been coupled with the SIMA agent for goal-directed navigation in generated environments. Its predecessors Genie 1 [10] and Genie 2 [20] charted the progression from static interactive platformers to foundation-scale environments. NVIDIA's Cosmos [51] and GR00T N1.6 [52] push foundation world models toward driving and humanoid robotics. Meta's V-JEPA 2 [44] demonstrates that prediction in latent space (without pixel generation) suffices for zero-shot planning.

Bush's Memex [11] was motivated by the inadequacy of hierarchical indexing for the “intricate web of trails” by which the mind operates. Engelbart [18] reframed this as augmentation. Nelson's Xanadu [48] envisioned persistent, typed, bidirectional links with full version history.

Four constructs are directly relevant. Typed links: Conklin [14] identified the typed link as hypertext's atomic unit; Halasz [29] argued for computed links from content analysis. Spatial hypertext: Marshall and Shipman [42, 55] showed that proximity and arrangement serve as implicit links. Nonlinear narrative: Bernstein [5] formalised structures including guard fields, cycles, and tangles. Reader agency: Landow [36] described the controlled-to-exploratory spectrum; Bolter [8] theorised the electronic “writing space.”

$mathcal {L}1$: Linear Chain. One successor per state. Hypertext analogue: axial text. $mathcal {L}2$: DAG. Branching, no cycles. Analogue: branching narrative. $mathcal {L}3$: General Graph. Cycles permitted. Analogue: tangled narrative. $mathcal {L}4$: Spatial Continuum. Continuous manifold with implicit proximity links. Analogue: spatial hypertext [42].

We type links as: causal (λc, action-conditioned), associative (λa, latent similarity), contrastive (λx, counterfactual comparison), hierarchical (λh, abstraction levels), and semantic (λs, linguistic affinity in language-grounded models). We propose that these types are identifiable through gradient attribution (λc), attention pattern analysis (λa), contrastive rollout comparison (λx), multi-scale architectural inspection (λh), and embedding similarity in language-aligned spaces (λs). We note that these identification methods remain to be validated empirically; we propose them as a concrete research programme rather than a demonstrated capability, and consider this operationalisation programme a contribution of the HWM framework.

$mathcal {T}1$: Sequential Simulation. One path, one successor per step. $mathcal {T}2$: Branching Exploration. Multiple simultaneous branches (e.g., MCTS [54]). $mathcal {T}3$: Parallel Global Prediction. Entire trajectory in one pass (diffusion [1], flow matching [40]). $mathcal {T}4$: Free Spatial Navigation. Unconstrained movement through continuous space (Genie 3 [21]).

We define α as the proportion of state transitions where the agent effectively controls the successor state. To ground this concretely:

Worked Example 1: DreamerV3 [28]. Over a 15-step imagination rollout, the agent selects 15 actions; the model generates 15 stochastic state samples. The agent controls which branch but the model controls branch content. α ≈ 0.50.

Worked Example 2: Sora [53]. A text prompt produces a 480-frame video with no frame-level control. One user decision, 479 model-determined transitions. α ≈ 0.002.

Worked Example 3: Genie 3 [21]. Continuous user input at 24 fps for one minute (∼ 1440 frames). The user issues directional/action inputs each frame; the model generates responsive but physically constrained frames. α ≈ 0.75.

This cluster comprises world models that generate single, linear trajectories with minimal or no agent intervention: autoregressive video generators (Sora [53]), driving scene generators (GAIA-1 [32]), and unconditional video models (SVD [7], Make-A-Video [56], CogVideo [31]). The linking topology is a simple chain ($mathcal {L}1$), all link types are causal (λc), and the agency coefficient approaches zero.

The critical question raised by Gupta [23], whether such systems qualify as world models at all, maps directly onto a hypertext distinction. In hypertext theory, a document with no links and no reader choice is not a hypertext; it is a linear text. By analogy, a generative model with no action-conditioning and no branching is not a world model in the navigational sense; it is a world narrator. We retain these systems in our taxonomy precisely to mark this boundary: the $mathcal {L}1/mathcal {T}1/alpha {approx }0$ corner is where “world model” shades into “world depiction.”

This reconceptualisation produces a concrete design recommendation: systems in this cluster can be upgraded to genuine world models by adding action-conditioned branching (moving from $mathcal {L}1$ to $mathcal {L}2$) and interactive control (increasing α). The Genie lineage demonstrates exactly this trajectory: Genie 1 [10] operated closer to $mathcal {L}1$ with limited interactivity; Genie 2 [20] added richer environment generation and preliminary interaction support; Genie 3 [21] achieves real-time spatial navigation at $mathcal {L}4$ with α ≈ 0.75.

This cluster includes model-based RL systems that explicitly construct and evaluate branching futures: the Dreamer family [25, 27, 28], MuZero [54], DIAMOND [1], IRIS [45], Δ -IRIS [46], TD-MPC2 [30], UniSim [65], and iVideoGPT [61]. The hypertextual structure is a DAG ($mathcal {L}2$): from each state node, the model generates multiple successor nodes corresponding to different actions, and the agent's traversal policy selects among branches based on value estimates.

DIAMOND is particularly interesting from the HWM perspective. By operating in pixel space rather than discrete latent space, it preserves the full “text” of each state node: no information is lost to tokenisation. DIAMOND's diffusion-based generation exemplifies what we term striated traversal: $mathcal {T}3$ (parallel global prediction via diffusion) within each frame, combined with $mathcal {T}1$ (sequential action selection) across frames. This hybrid modality, global within each node but sequential across nodes, has no clean analogue in prior hypertext theory.

The IRIS/Δ -IRIS family introduces a different hypertextual insight. By tokenising observations into discrete visual tokens, these models create an explicit vocabulary of world fragments, a finite set of reusable “lexia” from which world states are composed. The autoregressive Transformer then acts as a compositional linking engine, assembling lexia into trajectories according to learned grammatical rules. Δ -IRIS further creates edit-distance links whose semantics are defined by what changed between timesteps rather than what persists, a novel link type without precedent in classical hypertext theory.

This cluster encompasses world models that construct continuous spatial representations enabling free-form navigation: Genie 3 [21], navigation world models [4], GameNGen [58], GameGen-X [12], Oasis [16], Marble [60], and Gaussian splatting-based approaches.

The structural correspondence to spatial hypertext [42, 55] is striking. Just as spatial hypertext allows readers to discover structure through exploration of visually organised information space, spatial-immersive world models enable agents to discover environment structure through embodied exploration. Links are implicit, defined by spatial proximity and navigability rather than explicit graph edges. Organisation emerges from the geometry of the latent space rather than from authored connections.

Genie 3 is the paradigmatic example. Its consistency mechanisms, visual memory extending approximately one minute, frame-by-frame autoregressive generation conditioned on the full trajectory history, directly address what we term hypertextual friction (Section 8). The model's ability to maintain coherent environments during revisitation implements referential integrity across a navigational hypertext, one of Halasz's seven issues for next-generation hypermedia [29]. GameNGen and Oasis occupy hybrid $mathcal {L}2$/$mathcal {L}4$ positions: they create navigable spaces (spatial) but within constrained action vocabularies (partially branching), illustrating that real systems often straddle cluster boundaries.

The most recently emergent cluster couples language-based reasoning with dynamics prediction, creating world models that engage in a form of dialogue between model and agent. Joint MLLM-WM architectures [19], VLA models [52], WorldGPT [64], WorldCoder [13], Text2World [59], and RLVR-World [62] exemplify this paradigm.

The linking topology is a general graph ($mathcal {L}3$): the conversation between MLLM and world model creates cyclic dependencies (the MLLM's plan modifies the world model's predictions, which in turn inform the MLLM's replanning). Link types span all five categories: causal (action effects), associative (semantic similarity between task descriptions and states), contrastive (comparing plan alternatives), hierarchical (task decomposition into subtasks), and semantic (linguistic affinity between states).

From a hypertext perspective, this paradigm instantiates constructive hypertext [29]: the text (simulated world) is not pre-authored but is co-constructed through interaction. The HWM framework reveals a specific limitation: current MLLM-WM systems lack link type awareness. When an MLLM generates a plan and a world model validates it, neither system explicitly represents whether the connection between plan-step and world-state is causal, associative, or merely semantic. Introducing typed links at the MLLM-WM interface could enable more principled plan validation, for instance rejecting plans whose steps are connected only by semantic similarity rather than causal entailment.

SSMs as a Boundary Case. State-space models such as Mamba [22], when applied to world modelling, create exponentially decaying continuous-time links between states. This recency-weighted linking does not fit cleanly into our four topologies. Rather than introducing a premature “$mathcal {L}{1.5}$” category, we flag this as a limitation of the current framework and a direction for future refinement. The principled extension would require formalising how continuous-time decay kernels relate to the discrete link structures of classical hypertext.

The RSSM [25] decomposes latent state into a deterministic component ht = fθ(ht − 1, st − 1, at − 1) and a stochastic component st ∼ pθ(st|ht). In HWM terms, ht encodes the trail history, Bush's associative trail [11] rendered computational, while st introduces branching potential at each node. The model operates in the Branching-Explorative paradigm ($mathcal {L}2$, $mathcal {T}1$), with stochastic sampling creating a latent DAG of possible futures. The KL divergence term in the ELBO objective can be interpreted as quantifying the “semantic distance” between alternative narrative branches: high KL indicates strongly divergent futures, analogous to a hypertext choice point where different links lead to substantially different reading experiences.

PlaNet [26], which preceded the Dreamer series, used the same RSSM architecture but with cross-entropy planning (CEM) rather than actor-critic optimisation for action selection. In HWM terms, CEM implements a stochastic search over traversal paths: the agent samples many possible readings of the hypertextual future and selects the one with highest expected value, a form of “speed-reading” multiple narrative branches simultaneously.

IRIS [45] casts world modelling as sequence prediction over discrete visual tokens. A discrete autoencoder maps each frame into a short sequence of tokens from a learned codebook, and an autoregressive Transformer predicts future token sequences conditioned on past frames and actions. In HWM terms, the codebook functions as a finite vocabulary of world fragments, a set of reusable “lexia” from which world states are composed. The Transformer's self-attention mechanism then acts as a computed link generator [29]: each attention head dynamically creates weighted associations between all token pairs, implementing links computed from content analysis rather than explicit authoring.

Δ -IRIS [46] introduces a key innovation: rather than predicting full frames, it encodes only the stochastic delta between consecutive timesteps. In hypertext terms, this creates edit-distance links (λδ), connections whose semantics are defined by what changed rather than what persists. This is a novel link type with no direct precedent in classical hypertext theory, where links typically connect nodes based on content similarity or hierarchical relationship rather than differential change.

DIAMOND [1] applies denoising diffusion to world modelling, generating future frames through iterative refinement of Gaussian noise conditioned on the current state and action. The denoising process can be interpreted as gradual authoring: the world model progressively resolves ambiguity, moving from a maximally uncertain state (pure noise) to a determinate future. Early denoising steps establish coarse global structure, analogous to document outlines, while later steps fill in fine details, analogous to paragraph-level text. This multi-resolution structure introduces hierarchical links (λh) along a resolution axis.

DIAMOND's choice of the EDM (Elucidating the Design Space of Diffusion Models) framework over DDPM is significant from the HWM perspective: EDM provides substantially better stability over long imagination horizons, empirically reducing the per-step error amplification factor δt (formalised in Section 8). In hypertext terms, EDM reduces traversal friction, enabling longer navigational chains before disorientation sets in.

Flow matching [40], an emerging alternative to diffusion, replaces the noise schedule with learned optimal transport paths between distributions. In HWM terms, flow matching creates geodesic links: transitions that follow the shortest path through latent space, potentially reducing the “navigation cost” between states compared to diffusion's more circuitous denoising trajectory.

LeCun's JEPA proposal [37] and Meta's V-JEPA 2 [44] predict in representation space rather than pixel space. Rather than generating future observations, JEPA models predict the latent embedding of future states, avoiding the computational burden of pixel-level generation while preserving the information needed for planning.

In HWM terms, JEPA operates exclusively in a latent hypertext: a structure whose nodes are abstract representations rather than human-interpretable documents. Links are associative (λa), defined by embedding similarity rather than generative entailment. This architectural choice sacrifices readability (human inspection of individual state nodes) for navigational efficiency (fast traversal and planning in latent space). The trade-off recapitulates a tension in hypertext design between richly annotated nodes that support human comprehension and lightweight nodes optimised for machine processing. V-JEPA 2’s demonstrated capability for zero-shot planning suggests that navigational efficiency may suffice for many tasks, but the interpretability cost is real, and the HWM perspective makes this trade-off explicit.

The most recent architectural development couples world models with multimodal large language models, creating systems where natural language serves as both a control interface and a representation medium [19, 64]. WorldCoder [13] takes this further by building world models through code generation: the agent writes Python programs that simulate environment dynamics, creating an explicit, inspectable, and modifiable world model.

In HWM terms, language introduces semantic links (λs) connecting states that share conceptual affinity regardless of causal or spatial proximity. A state in which a cup is on a table is semantically linked to all states involving cups, tables, or placement actions, regardless of dynamic reachability. This creates a densely connected hypertext where navigation can proceed along causal, spatial, or semantic axes simultaneously, enabling the agent and model to engage in a turn-taking dialogue that co-constructs the experienced world. This directly mirrors the reader-text interaction theorised by Landow [36] and Bolter [8].

Driving world models [32, 50, 51] face a combinatorial linking challenge: the state space is a product of ego-vehicle, other-agent, and environmental states, creating an exponentially branching DAG. Current systems address this by conditioning on ego-actions while marginalising over other agents, implementing a perspectival constraint.

In HWM terms, this is structurally identical to Bernstein's guard field [5], a narrative device that reveals or conceals links based on the reader's accumulated context. The ego-vehicle is the “reader” whose perspective determines which state transitions are visible. Other agents’ futures are collapsed into probability distributions rather than navigable branches.

The HWM framework reveals a specific limitation: current perspectival constraints are fixed at design time (always ego-centric), whereas hypertext guard fields can be dynamically reconfigured. A driving world model that could temporarily shift perspective, “reading” the world from an oncoming vehicle's viewpoint to assess collision risk, or from a pedestrian's viewpoint to verify visibility, would occupy a different position on the α -axis and could improve safety validation. Concretely, this suggests an architecture with switchable conditioning: the same world model conditioned alternatively on ego-actions, adversary-actions, or joint actions, producing different traversal paths through the same state graph. This recommendation is invisible in architecture-centric surveys but follows directly from the hypertext concept of reader-reconfigurable guard fields.

Furthermore, the agency coefficient varies systematically with autonomy level. SAE Level 2 systems operate at α ≈ 0.3 (the human driver is primary “reader” with the model providing advisory “annotations”), Level 3 at α ≈ 0.5 (shared authorship with human override), and Level 4+ at α ≈ 0.7 (model and agent as collaborative co-authors of the trajectory). This progression mirrors hypertext's own evolution from author-controlled to reader-driven navigation.

Genie 3’s [21] core achievement, maintaining consistent environments for minutes of free exploration at 24 fps, is a direct implementation of referential integrity in generated hypertext. When a user navigates away from a location and returns, the model must ensure consistency despite the intervening trajectory being entirely user-driven. This is Halasz's referential integrity problem [29] transposed to generated content.

Genie 3’s visual memory (recall extending ∼ 1 minute) implements a bounded version: integrity degrades beyond the memory window, creating a “forgetting horizon” that is the temporal analogue of a hypertext system's link-update propagation boundary. DeepMind acknowledges this limitation: the model “can support a few minutes of continuous interaction, rather than extended hours” [21].

The HWM analysis suggests a specific improvement: hierarchical referential integrity, where coarse environmental structure (room layout, major landmarks, terrain topology) is maintained indefinitely via a persistent spatial graph ($mathcal {L}4$ at the coarse level), while fine details (object positions, lighting, surface textures) are maintained only within the memory window ($mathcal {L}1$ at the fine level). This mirrors multi-level caching in large-scale hypertext systems, where document structure is cached globally while content loads on demand. The practical architecture would couple a lightweight persistent 3D scene graph with Genie 3’s frame-level autoregressive generator, the scene graph providing structural anchors that prevent long-horizon drift.

GameNGen [58] illustrates a complementary approach: by training on a specific game (DOOM), it achieves stronger referential integrity within that domain because the game's fixed geometry provides implicit structural anchors. In HWM terms, DOOM's level geometry functions as a pre-authored spatial hypertext that constrains the model's generative freedom, trading agency for consistency.

Robotic manipulation tasks [52, 66] require world models that capture physical constraints, rigidity, friction, gravity, contact dynamics, as inviolable transition rules. In HWM terms, physical laws function as link axioms: structural constraints that no valid traversal may violate, analogous to integrity constraints in database-backed hypertext systems.

The emergence of foundation world models for robotics [43, 52] represents a shift from task-specific hypertext (a single document tailored to one reading purpose) to general-purpose hypertext (a docuverse from which multiple task-relevant reading paths can be extracted). NVIDIA's GR00T N1.6, for instance, trains a single world model across multiple humanoid platforms (YAM, AgiBot Genie-1, Unitree G1), then fine-tunes for specific tasks, analogous to a universal hypertext corpus from which domain-specific views are extracted through specialised link filters.

The HWM lens reveals that current robotic world models under-utilise contrastive links (λx). When a manipulation plan fails, agents typically replan from scratch rather than analysing counterfactuals: “what would have happened if the grasp angle had been 10 degrees different?” Explicitly maintaining contrastive links between executed and alternative trajectories would support more efficient learning from failure, a form of “what-if” navigation that hypertext supports naturally through bidirectional linking.

To demonstrate that the HWM framework generates falsifiable claims rather than merely descriptive labels, we derive a specific testable prediction for Genie 3 [21], which is publicly accessible to Google AI Ultra subscribers.

Genie 3 operates at $mathcal {L}4$ (spatial continuum) with bounded referential integrity: its visual memory extends approximately one minute. In HWM terms, this means the system maintains link validity (consistency between revisited nodes) only within a temporal window. The framework predicts that referential integrity should degrade faster when the user's trajectory has high cycle density (frequent revisitations of previously explored areas) than when the trajectory follows a linear exploration path. The reasoning is hypertextual: cyclic traversal stresses the bounded memory window by requiring simultaneous consistency across multiple revisited nodes, whereas linear traversal only requires forward consistency. Concretely, a user who walks in circles through a Genie 3 environment should observe visual inconsistencies earlier than a user who walks in a straight line, even if both sessions have the same total duration.

This prediction is testable with Genie 3’s public interface: generate two environments from the same prompt, explore one linearly and one cyclically, and measure the timestep at which visual inconsistencies first appear. If confirmed, it would validate the HWM framework's analytical utility; if refuted, it would reveal that Genie 3’s memory mechanism is more sophisticated than a simple bounded window, requiring refinement of the $mathcal {L}4$ characterisation. We propose this experiment as immediate future work.

To move beyond purely analytical claims, we present a preliminary empirical comparison of two Cluster 2 systems that share the same linking topology ($mathcal {L}2$) and agency coefficient (α ≈ 0.50) but differ in traversal modality: DreamerV3 [28] ($mathcal {T}1$, sequential) and DIAMOND [1] ($mathcal {T}1/mathcal {T}3$, striated). The HWM framework predicts that striated traversal should exhibit lower hypertextual friction than pure sequential traversal, because the diffusion-based per-frame generation ($mathcal {T}3$ within each step) acts as an error-correcting mechanism that pure autoregressive unrolling lacks.

We model error compounding using Equation 1 with per-step amplification factors estimated from published ablations: δt ≈ 0.035 for DreamerV3’s RSSM (based on reconstruction quality degradation beyond the default horizon H = 16 [28]) and δt ≈ 0.018 for DIAMOND's EDM framework (based on its reported stability advantage over DDPM, which exhibits δt ≈ 0.055 [1]). Table 3 reports cumulative error and the navigational radius, defined as the rollout horizon at which cumulative error reaches 2 × the initial error.

Table 3: Traversal Friction: Cumulative Error by Horizon

The results support the HWM prediction. DIAMOND-EDM's striated traversal extends the navigational radius by 86% (39 vs. 21 steps) compared to DreamerV3’s sequential traversal, despite both systems operating at the same $mathcal {L}2$ topology and α ≈ 0.50. At DreamerV3’s default imagination horizon of H = 16, DIAMOND-EDM accumulates 1.30 × less error. This lower friction correlates with downstream performance: DIAMOND achieves a mean human-normalized score (HNS) of 1.46 on Atari 100k, compared to 1.097 for DreamerV3, 1.046 for IRIS, and 1.266 for STORM [1], all of which use pure $mathcal {T}1$ traversal.

We emphasise that this analysis is model-based rather than directly experimental: the δt values are estimated from published ablations rather than measured through controlled experiments. Nevertheless, the consistency between the HWM-predicted friction advantage and the observed performance gap provides preliminary evidence that the traversal modality axis captures a real and consequential structural distinction.

The “lost in hyperspace” problem [14] has a direct computational analogue: compounding prediction error over long rollouts. Let ϵt denote prediction error at step t. In autoregressive world models, errors compound approximately as:

begin{equation} epsilon T le prod {t=1}^{T} (1 + delta t) cdot epsilon 0 end{equation} (1)

where δt is the per-step error amplification. This multiplicative bound, well-known in the model-based RL literature [28], is the computational analogue of navigational disorientation. DIAMOND [1] empirically demonstrates that the EDM diffusion framework significantly reduces δt compared to DDPM, improving long-horizon stability on Atari. Genie 3 extends coherence to minutes but acknowledges that “inaccuracies tend to accumulate over time” [21].

We propose latent navigation maps: compressed persistent representations providing global orientation independent of trajectory-level predictions. V-JEPA 2’s [44] latent space may provide the substrate. This directly imports hypertext's structural navigation aids (site maps, breadcrumbs) into world model design.

Perceptually convincing but dynamically incorrect futures constitute link integrity failure: visually valid transitions violating semantic constraints. DIAMOND [1] concretely illustrates this: its Atari world model sometimes permits double-jumps impossible in the actual game. Genie 3’s self-supervised physics may learn visually plausible but subtly incorrect dynamics.

We propose physical link axioms: formally specified transition constraints, analogous to database integrity constraints in hypertext systems, implementable via rejection sampling or physics-informed loss terms.

No current world model provides provenance for predictions. When a driving world model predicts that a pedestrian will stop, there is no mechanism to trace why. This is Nelson's Xanadu problem [48] for generated content. Interpretable world modelling with transitions carrying provenance metadata (attention attribution, uncertainty estimates) is critical for safety applications.

We propose six concrete directions at the intersection of world models and hypertext theory:

(1) Typed transition interfaces. Develop formal type systems for world model transitions that distinguish causal, associative, contrastive, hierarchical, and semantic links. This would enable richer querying of predicted state spaces (“show me all states causally reachable from here” vs. “show me all states semantically similar to here”) and more principled planning. The practical implementation might extend the attention mechanism with explicit link-type heads.

(2) Multi-resolution linking architectures. Design world models with explicit hierarchical structure: coarse-grained scene graphs linked to fine-grained physics simulations, enabling hierarchical referential integrity as discussed in Section 7. The practical substrate may be graph neural networks for coarse structure coupled with diffusion generators for fine detail.

(3) Bidirectional and lateral traversal. Enable backward navigation (“what states could have led here?”) and lateral navigation (“what parallel states exist at this timestep?”), extending the unidirectional forward rollout of current models. This is computationally expensive but would support richer counterfactual reasoning and debugging of model failures.

(4) Reader agency metrics. Develop standardised quantitative measures of agent control over world model traversal (operationalising α), enabling principled comparison across systems and informing the design of human-in-the-loop world modelling interfaces.

(5) Link provenance infrastructure. Build world models with built-in provenance tracking for transitions, enabling auditability and trust calibration for safety-critical domains. Lightweight approaches might store attention attribution summaries rather than full computation traces.

(6) Hypertextual evaluation benchmarks. Design evaluation protocols that assess not only prediction accuracy but also navigational richness: branching factor, cycle structure, semantic link diversity, and referential integrity duration. Current benchmarks such as Atari 100k and WorldSimBench [39] measure prediction quality but not structural richness. A “HyperBench” that evaluates world models on traversal coherence, counterfactual consistency, and provenance availability would directly operationalise the HWM framework.