World Narrative Model (WNM)

• World Narrative Model (WNM) is a framework that generates media by decoupling what to render from how to render using explicit 3D+T world representations.
• It employs a methodology that details scene geometry, object placement, and physical parameters to support counterfactual branching and controlled narrative evolution.
• WNM enhances planning and control by integrating structured state evolution, dedicated planning agents, and evaluation metrics to improve media synthesis.

World Narrative Model (WNM) most directly denotes a media-generation framework that decouples “what to render” from “how to render” by replacing end-to-end black-box pixel sampling with an explicit, editable, instance-level $4D$ $(3D+T)$ world representation (Chen et al., 30 Jun 2026). In adjacent research, closely related systems treat future observations, episodes, or narrative continuations as structured rollouts of latent world state rather than as unstructured outputs. This suggests a broader technical motif: explicit world state, temporally organized evolution, counterfactual branching, and controllable realization.

1. Definition, motivation, and intellectual scope

The named WNM formulation in video generation arises from a critique of mainstream video models as high-dimensional conditional samplers over pixels rather than models of the underlying physical world. In that formulation, video is treated as a $2D$ projection of an explicit $3D+T$ world, and the central design move is to build a controller that specifies scene geometry, object layouts, character or animal skeleton motion, trajectories, camera motion, and lighting at quantitative, physically meaningful granularity, while delegating final pixel synthesis to an existing video foundation model (Chen et al., 30 Jun 2026). The same paper frames industrial failure as a controllability problem: the “gacha” loop in which creators may require 20–50 generations, and for premium content more than 100 trials per shot, because prompt-level control does not expose exact object placement, joint angles, light positions, or lens parameters (Chen et al., 30 Jun 2026).

A broader world-model literature places this formulation inside a larger family of internal simulators. A survey account defines world models as internal representations used to understand and predict environmental dynamics, organized around an encoder, a dynamics module, and a decoder, and explicitly positions them as “embedded cognitive engines for autonomous agents” (Zhao et al., 31 May 2025). A critical essay pushes the point further, defining a world model by the transition $s' \sim p(s' \mid s,a)$ and arguing that its primary goal is to simulate “all actionable possibilities of the real world for purposeful reasoning and acting” (Xing et al., 7 Jul 2025). WR-Arena adopts the same transition formalism and argues that next-state prediction and perceptual realism are insufficient if a model cannot sustain long-horizon simulation, generate counterfactual rollouts, and support planning (Team et al., 26 Mar 2026).

Within that broader scope, WNM is not a single architecture. It names a direct controllable-video paradigm in one paper, but related work instantiates the same principle in embodied navigation, long-form serialized fiction, web-native persistent worlds, and graphical causal story models. The unifying commitment is that future output should be generated from an explicit or at least persistent world process, not only from local surface continuation.

2. Representational regimes

The direct WNM formulation for video generation represents the world as a structured physical narrative. Its controllable variables include scene geometry, object placements, character/animal skeleton motions, trajectories, camera motion, and lighting parameters; elsewhere the same representation is described as a $3D+T+Entity$ blueprint with $6$-DoF paths, focal length, depth-of-field, color temperature, beam angle, and optional plausibility constraints such as rigidity, collision, gravity, and light propagation (Chen et al., 30 Jun 2026). The point is not merely richer conditioning. The representation is intended to be fully editable, visualizable in pre-visualization form, and semantically aligned with professional filmmaking workflows.

A closely related decomposition appears in Web World Models, which factorizes state as

$S_t=(S_t^{\phi},S_t^{\psi}),$

with deterministic code updating the physics layer,

$S^{\phi}_{t+1}=f_{code}(S^{\phi}_t,a_t),$

and an LLM sampling the imagination layer,

$$S^{\psi}_{t+1}\sim \pi_{\theta}(\cdot \mid S^{\phi}_{t+1}).$$

That framework treats typed web interfaces as the latent state and insists that core rules and state transitions remain distinct from creative generation (Feng et al., 29 Dec 2025). This suggests a general WNM design pattern in which canon or mechanics are separated from narrative realization.

Shadow-Loom provides the most explicit narrative-state ontology. Its world state is

$(3D+T)$0

where nodes, edges, channels, world traits, and a branch tag $(3D+T)$1 together define a versioned graphical world model (Wilmot, 4 May 2026). The graph includes typed entities, event nodes, locations, objects, channels, and global traits; events carry both fabula time and syuzhet index, while beliefs store provenance through events and channels. This makes temporality, epistemics, and counterfactual branching first-class.

A more abstract formal narratology paper separates the underlying story world model $(3D+T)$2 from a reasoner’s fabula $(3D+T)$3, and then models communication as a mapping from Narrator-side world and fabula to Reader-side fabula and world reconstruction (Castricato et al., 2021). In that framework, the fabula is not the whole world but the set of propositions judged important for constructing an accurate story-world model. That distinction is highly consequential for WNM: it separates full latent world state from cognitively salient narrative abstraction.

3. Dynamics, memory, and counterfactual rollout

Embodied navigation work provides one operational lineage for WNM-like dynamics. Navigation World Models encode egocentric observations into latent states and model action-conditioned transition dynamics with

$(3D+T)$4

using a Conditional Diffusion Transformer and planning by simulating candidate trajectories, scoring them with perceptual similarity, and optionally optimizing them with the Cross-Entropy Method (Bar et al., 2024). In that setting, the “story” is a first-person rollout of what the world would look like under a candidate navigation action sequence.

Aerial Navigation World Model extends the same principle to $(3D+T)$5-DoF UAV control with

$(3D+T)$6

and augments latent diffusion with Future Frame Projection (FFP), a geometry-based prior that reprojects historical RGB-D observations into a future viewpoint (Zhang et al., 26 Dec 2025). Its planning loop ranks candidate trajectories by endpoint LPIPS to a target image. Quantitatively, on 2D trajectories at $(3D+T)$7s horizon, ANWM reports LPIPS $(3D+T)$8, DreamSim $(3D+T)$9, and FID $2D$0, against NWM’s $2D$1, $2D$2, and $2D$3; on 3D trajectories at $2D$4s it reports LPIPS $2D$5, DreamSim $2D$6, and FID $2D$7, against NWM’s $2D$8, $2D$9, and $3D+T$0 (Zhang et al., 26 Dec 2025). This is a narrow WNM, but it makes the action-conditioned visual narrative explicit.

For long-form fiction, N-VSSM treats serialized storytelling as a partially observed latent-state process and maintains an explicit narrative latent

$3D+T$1

updated once per scene, with posterior

$3D+T$2

where

$3D+T$3

Generation conditions on $3D+T$4 through a low-rank adapter inserted into an $3D+T$5B Mamba-2 decoder (Mann et al., 16 Jun 2026). On NarrativeWorldBench, the model holds plot-beat F1 $3D+T$6 across horizons $3D+T$7, whereas closed-frontier systems saturate in the band $3D+T$8 and collapse by about $3D+T$9 F1 at $s' \sim p(s' \mid s,a)$0 (Mann et al., 16 Jun 2026). Here the world narrative is not visual space but long-arc story continuity.

Shadow-Loom formalizes counterfactual branching directly. A factual world and a shadow branch coexist in the version DAG, Pearl’s ladder is implemented over the typed graph, and interventions use explicit edge surgery. Its abduction stage updates a trait by a precision-weighted blend of prior and evidence, while downstream propagation follows an “Impact $s' \sim p(s' \mid s,a)$1 Inertia” rule: $s' \sim p(s' \mid s,a)$2

$s' \sim p(s' \mid s,a)$3

This makes counterfactual narrative continuation a typed graph transformation rather than a free-form text rewrite (Wilmot, 4 May 2026).

4. Control, planning, and evaluation

The direct WNM system for video generation is explicitly agentic. It includes a Scene Layout Agent, Asset Generation Agent, Motion Planning Agent, and Camera and Lighting Agent, then turns an existing video foundation model into a “faithful neural shader” that renders from the world blueprint (Chen et al., 30 Jun 2026). The control surface is correspondingly explicit: a workflow canvas, entity node graph, narrative timeline, and four AI-augmented director consoles for scene and assets, character motion and trajectory, cinematography, and lighting and atmosphere. In evaluation, it reports overall GSB scores of 2.75 against a text-only baseline, 2.02 against an omni-reference professional baseline, and 1.72 for WNM+novice against the professional baseline; in the user study, novice users move from 18.6 to 5.1 trials per shot and from 74.2 min to 29.8 min, while professional users move from 9.4 to 3.2 trials and from 46.7 min to 24.1 min (Chen et al., 30 Jun 2026).

Navigation world models make the same control problem visible in a different domain. NWM exposes an energy

$s' \sim p(s' \mid s,a)$4

and plans by minimizing expected energy over imagined futures (Bar et al., 2024). OmniNWM generalizes this triad to state, action, and reward, jointly generating panoramic RGB, semantics, metric depth, and 3D occupancy, conditioning on normalized panoramic Plücker ray-maps, and deriving dense rewards from occupancy: $s' \sim p(s' \mid s,a)$5 Its long-horizon flexible forcing keeps FVD nearly flat from 23.63 at $s' \sim p(s' \mid s,a)$6 frames to 25.22 at $s' \sim p(s' \mid s,a)$7 frames, whereas the no-forcing variant degrades from 26.79 to 386.72 (Li et al., 21 Oct 2025). These systems are not narrative in the literary sense, but they operationalize WNM as controllable action-conditioned world rollout plus evaluable consequence.

WMNav shows a lighter-weight alternative in which prediction is externalized into explicit memory. PredictVLM estimates directional goal likelihood, projection writes those scores into a Curiosity Value Map, and the map is updated online by

$s' \sim p(s' \mid s,a)$8

This design is used to mitigate hallucination by grounding decisions in the difference between the world-model plan and subsequent observation, and it improves zero-shot object-goal navigation by +3.2% SR and +3.2% SPL on HM3D and +13.5% SR and +1.1% SPL on MP3D (Nie et al., 4 Mar 2025).

WR-Arena supplies a benchmark vocabulary for judging whether these systems are genuine world models. It evaluates Action Simulation Fidelity, Long-horizon Forecast, and Simulative Reasoning and Planning, and formalizes world-model rollout as

$s' \sim p(s' \mid s,a)$9

Its diagnosis is that perceptual quality alone is insufficient if a model cannot follow semantically meaningful multi-step instructions, sustain coherent futures, or support decision-making over alternatives (Team et al., 26 Mar 2026).

5. Narrative epistemics, reader states, and realism

A full WNM cannot be limited to physical control. Formal narrative theory introduces a second layer: what is true in the world versus what has been revealed and inferred. One model-theoretic account defines a story world model $3D+T+Entity$0, a reasoner’s fabula $3D+T+Entity$1, and a communication path from Narrator-side world and fabula to Reader-side fabula and reconstructed world (Castricato et al., 2021). Reader uncertainty is represented not as a single state but as a set of compatible worlds, and the paper proposes two coherence measures: Entropy of World Coherence (EWC) and Entropy of Transitional Coherence (ETC). That framework makes narrative understanding a trajectory of belief change rather than mere text continuation.

Shadow-Loom extends this epistemic line by turning the same world graph into a reader-state machine. With anchor $3D+T+Entity$2, it computes mystery, dramatic irony, suspense, and surprise as functionals over fabula time, syuzhet revelation, causal ancestry, and focal belief sets (Wilmot, 4 May 2026). Mystery is a hidden-cause ratio over revealed effects; dramatic irony is reader-character knowledge asymmetry; suspense is a balance-and-stakes score over unrevealed hope and threat; surprise is a saturated KL divergence between expected and actual final trait state. The important point is structural: affect is evaluated over world state and revelation structure, not inferred from tone alone.

Reduced but still informative narrative-world representations appear in network models of fiction. One such framework models a narrative as a dynamically unfolding character network enriched with chapter-level sentiment and topic-state vectors. For Les Misérables, it tracks nonlinear growth in nodes and edges, derives 21 Sequences from character-composition similarity, and interprets topical-state changes as interaction-induced state assimilation (Min et al., 2016). Another analysis of A Song of Ice and Fire separates story time from discourse time, finds that significant-death intervals are geometric in chapter order with

$3D+T+Entity$3

but power-law in story chronology with

$3D+T+Entity$4

and reports structural stability in a network of 2007 characters, with chapters stabilizing around 35 characters and major POV degree averaging 154.0 (Gessey-Jones et al., 2020). These works do not define WNM directly, but they show that narrative worlds require dual temporal layers, persistent social topology, and reader-cognitive scale.

6. Limitations, controversies, and research directions

Current WNM instantiations remain domain-narrow. The controllable-video WNM paper states that precision and robustness are still insufficient for full professional production, that motion/trajectory and object placement are handled separately, and that heavy reliance on LLM implicit reasoning for spatial understanding can be coarse (Chen et al., 30 Jun 2026). Navigation world models are likewise limited: NWM is explicitly a “narrow instance” focused on visual navigation, with short context and weak long-horizon consistency (Bar et al., 2024), while ANWM lacks language grounding in the model itself, persistent long-term memory, and abstract causal explanations such as “turning left reveals the plaza behind the tower” (Zhang et al., 26 Dec 2025).

Web World Models provide a complementary limitation and remedy. They argue that open-ended generative worlds are hard to control, debug, and keep consistent, and propose typed web interfaces, deterministic generation, and graceful degradation instead (Feng et al., 29 Dec 2025). This suggests that mature WNM systems may need a strict separation between code-defined rules or canon state and model-driven imagination, especially when persistence and revisitation matter. Shadow-Loom reaches a similar conclusion from the opposite direction: it uses LLMs only at the boundary and keeps identification, intervention, and counterfactual reasoning in typed code, but it also states that the graph is typed for narrative use rather than rigorous structural-causal identification and presents itself as a research artifact rather than a benchmarked NLP model (Wilmot, 4 May 2026).

Benchmark evidence remains sobering. WR-Arena reports that no model exceeds 65% on either transition smoothness or generation consistency, and explicitly states that current systems remain substantially below human-level hypothetical reasoning (Team et al., 26 Mar 2026). The critique literature argues that future world models should use hierarchical, mixed continuous/discrete representations, observation-grounded generative loss, and support imagined experience for reasoning and acting (Xing et al., 7 Jul 2025). This suggests that a fuller WNM would need to combine at least five ingredients: explicit state representation, long-horizon memory, counterfactual branching, reader- or agent-relative epistemics, and evaluation beyond perceptual fidelity.

One bibliographic caution is also notable. The record associated with “A System for Image Understanding using Sensemaking and Narrative” is, in the supplied material, actually a formatting template and “contains essentially no scientific content relevant to a World Narrative Model” (Battad et al., 2022). In a field where “world,” “narrative,” and “sensemaking” are increasingly cross-disciplinary, that caution is methodological as much as clerical: WNM synthesis depends on distinguishing conceptual alignment from genuine model content.