World Guild: Multi-Agent Scene Construction

• World Guild is a multi-agent framework that translates ambiguous natural language inputs into precise, executable scene descriptions for digital environments.
• It employs four specialized agents—Enricher, Manager, Critic, and Artist—in a sequential pipeline that refines, grounds, and validates scene layouts.
• The framework integrates error-correction datasets and procedural priors to ensure high fidelity, robust layout design, and efficient asset synthesis.

World Guild is a multi-agent framework within the World Craft system, designed to bridge the gap between ambiguous, natural language user instructions and precise, executable structured content suitable for game-engine environments. As a core subsystem of World Craft, World Guild orchestrates four specialized agents—Enricher, Manager, Critic, and Artist—in a sequential, cooperative pipeline that progressively interprets, grounds, refines, and visualizes scene descriptions for interactive digital environments. This framework systematically mitigates the “semantic gap” inherent in mapping narrative intent to the stringent spatial and syntactic requirements of real-time navigation and rendering backends (Sun et al., 14 Jan 2026).

1. System Overview and Role in World Craft

World Craft comprises two principal subsystems: World Scaffold and World Guild. World Scaffold defines a unified protocol and executable runtime to instantiate structured scene data into interactive, navigable, and visually rendered game environments. World Guild serves as the translation and synthesis module, converting ambiguous, often under- or over-specified user narratives into a level of formalism compatible with World Scaffold.

World Guild decomposes the end-to-end problem of scene construction into four stages, mapped to its distinct agents. These agents perform semantic enrichment, layout grounding, iterative critique and correction, and asset synthesis. This multi-agent design contrasts with single-step LLM-based or code-agent pipelines, targeting improved modularity, controllability, and error recovery.

2. Multi-Agent Architecture and Pipeline

The World Guild pipeline consists of the following agentic components and data transformations:

Agent	Input	Output	Core Function
Enricher	$\mathcal{I}$ (instruction)	$\mathcal{Z}$ (abstract topology, connectivity, zones)	Semantic sketching & disambiguation
Manager	$\mathcal{Z}$	$\mathcal{G}_0 = (M, A, L, P)$	Layout formalization & grounding
Critic	$\mathcal{G}_t$	$\mathcal{C}_t$ (correction proposals)	Rule-based & VLM-based validation
Artist	Final $\mathcal{G}$	Tiles for assets	Style-consistent asset rendering

• Enricher receives natural language input and generates an abstract, coordinate-free layout description, structuring user intent into coherent functional and topological zones.
• Manager grounds these into a structured format, instantiating metadata ($M$), an asset set ($A$), explicit layout coordinates ($L$), and interaction properties ($\mathcal{Z}$0).
• Critic iteratively applies rule-based (e.g., collision, connectivity) and vision-language semantic checks to identify violations or inconsistencies, providing targeted corrective feedback ($\mathcal{Z}$1) until a satisfactory scene state is reached or a maximum number of refinement steps ($\mathcal{Z}$2) is exceeded.
• Artist retrieves style- and dimension-matched reference assets and synthesizes visual tiles, which are subsequently assembled within World Scaffold into a final, playable scene.

This pipeline enables robust handling of ambiguous or error-prone user inputs by decomposing the horizontal complexity typically seen in one-shot LLM-based systems, and ensures higher overall stability and fidelity.

3. Error-Correction Dataset and Procedural Priors

A high-quality, reverse-engineered error-correction dataset underpins World Guild’s robustness and capability for spatial reasoning. Scenario domains cover Real-World, Literature, Film, and TRPG; 125 seeds per domain are augmented by five stylistic variants, yielding 2,000 training descriptions.

Golden layouts are constructed by an LLM-driven procedural generator informed by RPLAN-inspired architectural priors (e.g., orthogonality, central connectivity), with physical placement validation to ensure no object collisions. Teacher models and human experts then review and refine these layouts, producing $\mathcal{Z}$3, the standard for correctness.

Controlled corruption is introduced at four perturbation levels using a “Chaos Monkey” approach, randomly applying 2–15 errors per scene to generate $\mathcal{Z}$4 tuples. Datasets for model training are structured as follows:

• $\mathcal{Z}$5: $\mathcal{Z}$6
• $\mathcal{Z}$7: $\mathcal{Z}$8 with $\mathcal{Z}$9

Procedural rules derived from RPLAN (orthogonality, periphery-to-center layout, path optimization) are embedded both in gold layouts and correction trajectories to impart robust architectural priors.

4. Learning Objectives, Model Fine-Tuning, and Core Algorithms

The agent policies and models within World Guild are fine-tuned via a two-stage supervised objective:

• Stage 1 (Enricher, $\mathcal{Z}$0): Maximize $\mathcal{Z}$1 over $\mathcal{Z}$2:

$\mathcal{Z}$3

• Stage 2 (Manager, $\mathcal{Z}$4): Unify generation and error correction using $\mathcal{Z}$5:

$\mathcal{Z}$6

with $\mathcal{Z}$7.

The generative objective decomposes as $\mathcal{Z}$8, formally reflecting the intermediate role of $\mathcal{Z}$9 in the pipeline.

The critic-manager refinement loop is given by:

$\mathcal{G}_0 = (M, A, L, P)$0

Asset retrieval employs a penalty term $\mathcal{G}_0 = (M, A, L, P)$1 to enforce size consistency.

5. Evaluation Protocols and Performance Benchmarks

World Guild’s performance is quantized using multidimensional metrics:

• Layout Rationality: Collision-Free Rate (CFR, higher is better), Room Connectivity Score (RCS, higher is better), Object Placement Score (OPS, lower is better)
• Element Design: Component Existence Rate (CER, higher is better), Object Volume Density (OVD, higher is better), Physical Attribute Consistency (PAC, lower is better)
• Intent Alignment: VSA-C (CLIP-based), VSA-V (VLM-based), both higher is better

Key comparative findings:

System	CFR	RCS	OPS	CER	PAC	Time-to-Run (min)	HWR
Direct Gen (Gemini-3-Pro few-shot)	0.80	0.59	5.56
+ Critic (ablation)	0.84	0.67	4.99
Enricher+Manager+Critic (World Guild)	0.94	0.88	3.03
Two-stage + Critic (with correction)				0.99	3.64
Cursor (code agent)						15.4	0.23
Antigravity (code agent)						9.8	0.35
World Guild (fully automated)						4.3	0.92

Pearson’s $\mathcal{G}_0 = (M, A, L, P)$2 across evaluation dimensions and Fleiss’s $\mathcal{G}_0 = (M, A, L, P)$3 denote high inter-rater reliability for human annotation.

6. Illustrative Processing Examples

Sample processing illustrates the operation of World Guild’s pipeline:

• Input: “A luminous library–café with a hidden alchemist’s lab behind the bookcase.”
• Enricher Output ($\mathcal{G}_0 = (M, A, L, P)$4):
  • Zones: Library (front), Café (center), Lab (hidden, back)
  • Connectivity: café$\mathcal{G}_0 = (M, A, L, P)$5library via archway; hidden exit behind bookcase
• Manager Output ($\mathcal{G}_0 = (M, A, L, P)$6):
  • Metadata (grid size $\mathcal{G}_0 = (M, A, L, P)$7, style pixel art)
  • Asset list (bookshelf, table, stove, cauldron, door_hidden)
  • Layout: coordinate-precise item placement
  • Properties: collision and navigation tags
• Critic Loop: Detects initial blocked path, proposes movement, validated in next round.
• Artist: Retrieves and synthesizes pixel-art “wooden bookshelf” tile, then World Scaffold assembles the complete, functional, and navigable scene.

7. Significance and Limitations

World Guild exemplifies the potential of explicit multi-agent decomposition for complex grounding of free-form user narratives into executable environments. The approach outperforms baseline LLM-only and commercial code-agent pipelines in layout quality, design fidelity, intent alignment, and automation efficiency. A plausible implication is the scalability of this model to other domains requiring highly structured synthesis from unstructured input; however, performance remains gated by training data diversity and the efficacy of embedded procedural priors. Further developments may be expected in scaling multi-agent model complexity and dataset breadth within similar frameworks (Sun et al., 14 Jan 2026).