MindJourney: Test-Time Scaling with World Models for Spatial Reasoning

arXiv 25.7 2025 62 method

TLDR

MindJourney couples VLMs with a controllable video diffusion world model for test-time scaling, improving spatial reasoning without fine-tuning.

Reasoning

The paper presents a novel test-time scaling framework that integrates a world model with VLMs, achieving notable performance gains on a spatial reasoning benchmark without fine-tuning. However, it lacks real-world deployment evidence and may face computational challenges due to video diffusion.

Read-first score

Read-first score 62, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 55.

Recency 8%

86.7

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

Topical relevance 42%

78.6

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

Methodology quality 25%

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

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 434.

Keyword Scores

world model

generative world model

interactive world model

video world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Test-time scaling framework that couples a VLM with a controllable world model based on video diffusion for spatial reasoning
Iterative process where VLM sketches a camera trajectory, world model synthesizes corresponding views, and VLM reasons over multi-view evidence
No fine-tuning required; plug-and-play integration with existing VLMs
Improves upon test-time inference VLMs trained through reinforcement learning

Methodology

MindJourney couples a vision-language model (VLM) with a controllable world model based on video diffusion. At test time, the VLM iteratively sketches a concise camera trajectory, the world model synthesizes the corresponding view at each step, and the VLM reasons over the gathered multi-view evidence. The framework is evaluated on the SAT spatial reasoning benchmark without any fine-tuning.

Key Results

MindJourney achieves an average 7.7% performance boost on the SAT benchmark without fine-tuning, and also improves upon test-time inference VLMs trained through reinforcement learning.

Resources

Paper Project