MineWorld: a Real-Time and Open-Source Interactive World Model on Minecraft

arXiv 25.4 2025 77 method, system, application

TLDR

MineWorld is a real-time interactive world model for Minecraft using a visual-action autoregressive Transformer with parallel decoding, outperforming diffusion-based models.

Reasoning

Strengths include a novel parallel decoding algorithm enabling real-time interaction and new metrics for evaluating action following. Weaknesses are the focus solely on Minecraft without discussion of generalization to other environments.

Read-first score

Read-first score 77, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 58.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%

82.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=code,github

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=evaluation,metric

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 17.

Keyword Scores

world model

interactive world model

generative world model

world dynamics prediction

world simulator

video world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Real-time interactive world model on Minecraft using a visual-action autoregressive Transformer
Novel parallel decoding algorithm that predicts spatial redundant tokens in each frame simultaneously, enabling 4-7 frames per second
New evaluation metrics that assess both visual quality and action following capacity for world models

Methodology

MineWorld uses a visual-action autoregressive Transformer that takes paired game scenes and actions as input, tokenized into discrete token ids via an image tokenizer and an action tokenizer, with the input formed by concatenating interleaved token ids. The model is trained with next token prediction to learn game state representations and state-action conditioning. During inference, a parallel decoding algorithm predicts spatial redundant tokens in each frame at the same time to achieve real-time interaction.

Key Results

MineWorld generates 4 to 7 frames per second, enabling real-time interaction, and significantly outperforms state-of-the-art open-source diffusion-based world models on the proposed evaluation metrics.

Resources

Paper Code