FantasyWorld: Persistent Virtual World Modeling

Updated 3 July 2026

FantasyWorld is a research field exploring methods for generating, representing, and interacting with persistent, narrative-rich virtual worlds.
It employs dual-state formalism that combines deterministic, code-driven physics with stochastic, LLM-generated narrative content to maintain logical coherence.
Innovative pipelines and human-AI co-creativity frameworks are advancing methodologies for procedurally generated, explorable, and visually immersive environments.

FantasyWorld, in contemporary research, refers to both the technical challenge and the suite of methodologies for generating, representing, and interacting with persistent, explorable, and logically consistent virtual worlds—particularly those with fantastical, narrative-driven, or game-like properties. It encompasses systems ranging from neural world-building pipelines to LLM-enhanced game engines, graph-structured knowledge corpora, and co-creative AI assistants, unified by the goal of supporting open-ended, persistent, and semantically rich fantasy environments suitable for human-AI interaction and narrative experience.

1. Formal Models and Architectural Principles

A central architectural innovation in the modern FantasyWorld paradigm is the layered world-state formalism, as exemplified by Web World Models (WWMs) (Feng et al., 29 Dec 2025). Here, world state is decomposed as

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

where $S^{\phi}$ encodes deterministic, code-executed "physics" (location, inventory, resource constraints, permissions, etc.), and $S^{\psi}$ encodes the stochastic, LLM-generated "imagination" (narrative context, lore, dialog, high-level perceptual phenomena). This dual-state formulation supports strict separation of enforceable invariants from generative narrative content. Deterministic update functions propagate world rules and events: $S^{\phi}_{t+1} = f_{code}(S^{\phi}_t, a_t)$ while the generative layer synthesizes contextual outputs conditioned on the new symbolic state: $S^{\psi}_{t+1} \sim \pi_\theta(\cdot \mid S^{\phi}_{t+1})$ Typed web interfaces (JSON, TypeScript) act as schemas and contracts, binding LLM generations to executable and coherent internal state, thereby enforcing logical validity and preventing structural hallucinations. Deterministic procedural generation (hashing locations into seeds) achieves persistent yet unbounded world growth, with state revisitation guaranteed by seed functions, obviating the need to materialize infinite universes in explicit databases.

PDVA (Plan-Diff-Validate-Apply) pipelines in orchestrated reality architectures extend this principle by stratifying action processing into planning, proposing narrative deltas, schema/rule checking, and atomic world state commits with content hashing. The world becomes an auditable, canonical object governed by Markovian transitions over structured, schema-valid JSON hierarchies (Huang et al., 14 Jun 2026). Parameterized actions and observation functions formalize the player-world interface as a parameterized-action POMDP: $\mathcal{M} = (\mathcal{S}, \mathcal{A}, \Omega, O, F, \rho_0)$ where all updates flow through a singleton orchestration agent analogous to the Game Master, with persistent context injection ensuring durable memory.

2. Data Resources and Knowledge Representation

High-fidelity fantasy worlds require dense, semantic knowledge bases. The Forgotten Realms Wiki (FRW) corpus is representative: an 11-view, 45,000-article dataset constructed from the largest D&D wiki, providing raw text, JSON-indexed content, infoboxes, directed article graphs, link hierarchies (first-link and all-links), word and document embeddings, and Poincaré embeddings of hierarchical page trees (Peiris et al., 2022). Such resources support:

NLP and graph-based similarity evaluation (Wu & Palmer, Jiang–Conrath, Poincaré, Word2Vec/Doc2Vec vectors)
Dense entity representations and lookup by canonical titles
Named entity classification into ontology categories using infoboxes (99.3% accuracy on generated text categorization tasks).

The corpus enables not only generation but fine-grained lore and continuity evaluations, supporting conditional text generation pipelines that balance data-driven and rule-based constraints.

3. Generative World Creation and Interaction

Procedural world generation from text (as in the LIGHT and FIREBALL settings) frames worldbuilding as a compositional learning task. Models are trained to assemble maps, place characters and objects, learn container-object affordances, and generate new content (locations, personas, descriptions, affordances) using transformer-based seq2seq and ranking architectures (Fan et al., 2019).

World graphs are constructed using ranked candidate selection for adjacency; filler locations ensure map diversity and coherence.
Generated descriptions, personas, and backgrounds are composed from open vocabulary and guided by pretrained creative corpora (e.g., Reddit pretraining for imaginative range).
Assistive interfaces allow human designers to co-iterate with the model, rapidly prototyping worlds via autocomplete-style suggestions and layered editing, as reflected in increased creative speed and diversity.

More recent VR systems (e.g., GenLARP) extend this to immersive, live-action experiences, mapping narrative intent through semantic extraction, structured prompts, procedural 3D scene construction, and live agent-driven characters with persistent memory, emotion, and motivation state (Yu et al., 16 Oct 2025).

4. Visual and Spatial Synthesis Pipelines

Multi-modal and spatial syntheses address the limitations of text-only or static-image-based worldbuilding, focusing on interaction, editability, and coherence.

NeoWorld employs an object-centric hybrid 2D/2.5D/3D pipeline, progressively "unfolding" the world from a single input image as user exploration advances. Foreground objects are lifted to full 3D via instance-aware segmentation and image-to-3D translation (Amodal3R); alignment, quality checks, and LLM-driven control/proxy for object selection and animation enable efficient, consistent, and physics-aware manipulation (Zhao et al., 29 Sep 2025).
World Craft introduces a multi-agent staged approach, mapping user textual description $\mathcal{I}$ to executable scene graphs $\mathcal{G}=(M,A,L,P)$ . Agents for semantic enrichment, constrained layout, critique/correction, and asset synthesis iteratively refine environments, using reverse-engineered correction data to improve spatial robustness. The system outperforms general code agents and LLMs in scene construction, intent alignment, and human preference, covering layout physicality, asset coherence, and correction stability (Sun et al., 14 Jan 2026).
WorldSmith and related tools support iterative, compositional visual worldbuilding, fusing text prompts, sketching, regional masks, layered edits, and global tile blending. This interactive, hierarchical interface supports multi-scale design from micro assets to macro world maps, aligns closely with fantasy world-creation workflows, and enables asset reuse, region-aware prompting, and branching designs (Dang et al., 2023).
For panoramic world "unrolling," equirectangular projections and inpainting-style restoration enable immersive, navigable 3D scenes from single panoramas, with geometric and perceptual correction ensuring local and global spatial consistency (Tikhonov et al., 2023).

5. AI Co-Creativity, Assistance, and Human-in-the-Loop Systems

A key thread is the design of AI systems that augment, rather than supplant, human creative agency in FantasyWorld construction and narration.

CALYPSO exemplifies DM-assistive LLM platforms tailored for table-top roleplay. It structures assistance into context summarization, focused brainstorming, and open-domain generation that is curated by the DM, balancing thematic commonsense, inspiration, and real-time fidelity without overstepping narrative control (Zhu et al., 2023).
FIREBALL provides stateful, game-mechanical representations of play-by-post D&D, aligning natural language with actual engine state transitions. LLMs finetuned with structured game state ( $S$ ) markedly outperform baselines in both command generation ("utterance to Avrae command") and narration ("state to description") (Zhu et al., 2023).
Orchestrated reality approaches formalize the world as a singleton-owned object, with memory injection and content-hashed validation, ensuring persistent NPC agency, consequence durability, and auditing. Interaction models are parameterized (intent, structured parameter bundle) and world updates are strictly validated for logical and rule compliance (Huang et al., 14 Jun 2026).

6. Limitations, Open Problems, and Future Directions

Although progress is rapid, several core limitations are recurrent:

Persistent determinism is nontrivial: stochasticity in generative outputs requires rigorous request/response auditing for full replay.
Narrative drift and schema coverage: features not canonically encoded (visual traits, subtle factions) can drift across sessions, as only schema-validated properties are enforced.
Multi-agent scalability: concurrency bottlenecks (cost and latency) can impact the experience as more NPCs or autonomous agents are introduced.
World coherence in compositional visual systems: maintaining global, cross-region aesthetic and spatial consistency remains an open research problem, especially under multi-modal edits and layer recombination.
Human-AI collaboration: best practices for creative agency, frictionless interaction, and synchronous co-development are still being optimized in real scenarios.

Empirical evaluations increasingly emphasize human preference, intent alignment, and system stability as primary outcome measures, with grounded benchmarks on scene correctness, narrative fidelity, and domain-specific logic.

7. Representative Systems and Applications

System/Paper	Core Modality/Contribution	arXiv ID
Web World Models (WWM)	Hybrid code+LLM world state, persistence	(Feng et al., 29 Dec 2025)
Orchestrated Reality	Canonical JSON world, PDVA pipeline, POMDP	(Huang et al., 14 Jun 2026)
NeoWorld	Progressive hybrid 2D-3D world unfolding	(Zhao et al., 29 Sep 2025)
World Craft	Multi-agent text-to-executable-symbolic	(Sun et al., 14 Jan 2026)
WorldSmith	Iterative, sketch/region-based visual build	(Dang et al., 2023)
GenLARP	Personalized VR LARP, multi-agent roles	(Yu et al., 16 Oct 2025)
FRW Dataset	Multi-view D&D knowledge base for modeling	(Peiris et al., 2022)
FIREBALL	State-grounded D&D discord corpus	(Zhu et al., 2023)
CALYPSO	Live LLM DM assistant	(Zhu et al., 2023)
LIGHT (World Generation)	Hierarchical, compositional text world-gen	(Fan et al., 2019)

These systems demonstrate the increasing integration of deterministic logical cores with generative, narrative overlays; the rise of structured, knowledge-aware corpora for lore and reasoning; and the evolution toward persistent, interactive, and hierarchically organized fantasy spaces that maintain logical, visual, and narrative coherence across user-driven exploration.

A plausible implication is that the convergence of explicit world state, schema-enforced imagination, multi-agent reasoning, and co-creative user interfaces is transforming the concept of FantasyWorld from a purely notional or textual artifact into a rigorously modelable, extensible, and persistently explorable environment, with strong methodological roots in both symbolic and neural paradigms. These developments position FantasyWorld research at the intersection of generative AI, game AI, knowledge engineering, HCI, and formal modeling, with far-reaching implications for virtual worlds, game design, simulation-based storytelling, and interactive narrative AI.