Multiverse: Language-Conditioned Multi-Game Level Blending via Shared Representation

In-Chang Baek; Jiyun Jung; Geum-Hwan Hwang; Sung-Hyun Kim; Kyung-Joong Kim

cs.AIMar 25, 2026

Multiverse: Language-Conditioned Multi-Game Level Blending via Shared Representation

Authors: In-Chang Baek, Jiyun Jung, Geum-Hwan Hwang, Sung-Hyun Kim, Kyung-Joong Kim

Organizations: Gwangju Institute of Science and Technology (GIST), South Korea · Dongguk University, South Korea

Abstract

Text-to-level generation aims to translate natural language descriptions into structured game levels, enabling intuitive control over procedural content generation. While prior text-to-level generators are typically limited to a single game domain, extending language-conditioned generation to multiple games requires learning representations that capture structural relationships across domains. We propose Multiverse, a language-conditioned multi-game level generator that enables cross-game level blending through textual specifications. The model learns a shared latent space aligning textual instructions and level structures, while a threshold-based multi-positive contrastive supervision links semantically related levels across games. This representation allows language to guide which structural characteristics should be preserved when combining content from different games, enabling controllable blending through latent interpolation and zero-shot generation from compositional textual prompts. Experiments show that the learned representation supports controllable cross-game level blending and significantly improves blending quality within the same game genre, while providing a unified representation for language-conditioned multi-game content generation.

Figures & tables

Fig. 1: Shared Latent Space Visualization. The proposed inter-game contrastive learning learns 128-dimensional representations for levels and instructions across multiple games, which are projected into three dimensions using t-SNE. Embeddings from different games are aligned within a unified latent space. The blended sample is generated by an interpolation between Dungeon and Lode Runner level embeddings, demonstrating cross-game continuity in the shared space.

Fig. 2: Multiverse Training Pipeline and Level Blending Mechanism. (a) A shared multi-game latent space is learned using multi-positive contrastive learning over text–level pairs. (b) The learned level embedding is used as a conditional vector for a VAE-based generator. For single-game generation, the encoded embedding directly conditions the decoder. For multi-game level blending, embeddings from different games are interpolated in the latent space, and the interpolated vector is provided to the decoder to synthesize blended levels.

Dataset	Single-game	Multi-game
Game	Single CL	Single CL	Multiverse
Dungeon (G1)	0.794 $\pm$ 0.004	0.730 $\pm$ 0.003	0.728 $\pm$ 0.005
Zelda (G2)	0.767 $\pm$ 0.009	0.687 $\pm$ 0.016	0.719 $\pm$ 0.020
SuperMario (G3)	0.623 $\pm$ 0.010	0.593 $\pm$ 0.008	0.599 $\pm$ 0.007
Lode Runner (G4)	0.792 $\pm$ 0.005	0.800 $\pm$ 0.003	0.799 $\pm$ 0.005
Overall	0.744 $\pm$ 0.082	0.703 $\pm$ 0.087	0.711 $\pm$ 0.083

TABLE I: Single-game generation VITScore performance under single-game instructions.

Panel of Figure 4 — Fig. 3: Embedding Interpolation-based Level Blending. Levels generated by interpolating between the embeddings of Super Mario (A) and Lode Runner (B) across different mixing ratios.

Group	Ratio A	Ratio B	ViTScore A	ViTScore B
Intra-genre (G1–G2 G3–G4)	0.00	1.00	0.392 $\pm$ 0.005	0.664 $\pm$ 0.011
	0.25	0.75	0.379 $\pm$ 0.006	0.650 $\pm$ 0.015
	0.50	0.50	0.501 $\pm$ 0.008	0.493 $\pm$ 0.011
	0.75	0.25	0.645 $\pm$ 0.012	0.394 $\pm$ 0.006
	1.00	0.00	0.671 $\pm$ 0.011	0.371 $\pm$ 0.004
Inter-genre (G1/G2– G3/G4)	0.00	1.00	0.411 $\pm$ 0.004	0.664 $\pm$ 0.010

TABLE II: Multiverse embedding-based blending generation results across mixing ratios.

Group	Method	Correlation	$\boldsymbol{\mathrm{WGM}_{\text{balance}}}$	$\boldsymbol{\mathrm{WGM}_{\text{bias}}}$
Intra-genre	Single CL	0.558 $\pm$ 0.017	0.477 $\pm$ 0.004	0.549 $\pm$ 0.006
	-Abstraction	0.548 $\pm$ 0.016	0.473 $\pm$ 0.004 *	0.546 $\pm$ 0.003
	-Threshold	0.556 $\pm$ 0.012	0.479 $\pm$ 0.004	0.548 $\pm$ 0.006
	Multiverse	0.580 $\pm$ 0.019 *	0.481 $\pm$ 0.009	0.559 $\pm$ 0.008 *
Inter-genre	Single CL	0.482 $\pm$ 0.012	0.493 $\pm$ 0.008	0.563 $\pm$ 0.004
	-Abstraction	0.469 $\pm$ 0.012 *	0.492 $\pm$ 0.007	0.558 $\pm$ 0.004 *

TABLE III: Comparison of generation models for level blending. Correlation evaluates controllability of embedding interpolation. Generation quality is measured using weighted geometric mean scores.

Panel of Figure 7 — Fig. 4: Multi-game Text Instruction-based Level Blending. Levels generated by composite text instruction between the embeddings of Super Mario (A) and Lode Runner (B) across different instruction combination strategies.

Fig. 5: Cosine Similarity Distribution of Composite Instructions. 2D kernel density estimation visualization of cosine similarity between composite instructions and the original instructions from game A and game B. Balanced combinations remain relatively symmetric, while biased combinations shift toward the axis corresponding to the base instruction.

Strategy	Composite Instruction
Text A	square floor room with central block cluster and surrounding wall
Text B	spiral wall path with bats clustered on right side
Concat	square floor room with central block cluster and surrounding wall spiral wall path with bats clustered on right side
Mix	A square-floored room is enclosed by a continuous wall that folds into a spiral path around a central cluster of blocks , and along the right side of the winding wall the bats gather in a dense, shadowy group .
A-base	based on square floor room with central block cluster and surrounding wall , featuring spiral wall path with bats clustered on right side .
B-base	based on spiral wall path with bats clustered on right side , featuring a square floor room with a central block cluster and surrounding wall .

TABLE IV: Example of composite instruction construction. Text from Game A ( The Legend of Zelda ) and Game B ( Dungeon ) are highlighted.

Procedural Content Generation

Procedural content generation creates game levels, 3D worlds, assets, stories, and simulation environments through algorithms and learned models.

See all 66 papers

Code2Games: Enabling Coding Agents for Gaming World Generation
Oct 4, 2026Wei Wu, Ziyang Xu, Zeyu Zhang +2Agentic Code Generation 3D Scene Generation
School of Computer Science, Peking University · La Trobe University
GAMEGO: Training Game-Dev Agents with Synthetic Trajectories Anchored in Real-World Assets
Oct 2, 2026Haoyue Yang, Jingyao Li, Zhengfan Wu +3Agentic Code Generation Synthetic Data Generation for Language Models
Institute of Automation, Chinese Academy of Sciences · Baidu Inc., Beijing, China
MultiGen: Level-Design for Editable Multiplayer Worlds in Diffusion Game Engines
Mar 3, 2026Ryan Po, David Junhao Zhang, Amir Hertz +3Interactive World Models Video World Models
Stanford University · Google

Oct 4, 2026J/K move · Enter open · S save

Code2Games: Enabling Coding Agents for Gaming World Generation

Wei Wu, Ziyang Xu, Zeyu Zhang, Yang Zhao+1

School of Computer Science, Peking University · La Trobe University

Agentic Code Generation 3D Scene Generation Procedural Content Generation

Open paper AskPDF arXiv

Generating a high-quality gaming world from a natural-language game intent requires joint reasoning about scene structure, spatial layout, gameplay objectives, interactive entities, and executable gameplay logic. Existing coding agents can generate individual assets, scenes, or scripts, but often struggle to maintain consistency across these components. We propose Code2Games, an agentic framework that builds a structured gaming world upon a base Blender world generated from the same game intent. Code2Games coordinates scene analysis, gameplay planning, constrained gaming-world generation, and gaming-engine customization through a shared scene-gameplay representation with persistent element correspondence. After world generation, Code2Games adapts the generated world to Unreal Engine 5 and employs an execution-guided reconstruction process that uses compilation diagnostics, runtime feedback, and gameplay test results to resolve inconsistencies arising during engine adaptation. To systematically evaluate gaming-world generation, we introduce the GameCode4D benchmark, which comprises ten fixed game prompts spanning different levels of scene and gameplay complexity. We evaluate the generated results across four dimensions: visual quality, interactive fidelity, multimodal artifact quality, and playable-game quality. Experiments demonstrate that, compared with direct gaming-world generation by coding agents and existing baseline methods, Code2Games consistently improves the visual quality and interactive fidelity of generated gaming worlds, as well as the quality of the resulting games after engine adaptation.

Abstract

Procedural Content Generation

Explore similar work

Code2Games: Enabling Coding Agents for Gaming World Generation

GAMEGO: Training Game-Dev Agents with Synthetic Trajectories Anchored in Real-World Assets

MultiGen: Level-Design for Editable Multiplayer Worlds in Diffusion Game Engines