MindJourney: Test-Time Scaling with World Models for Spatial Reasoning @MicrosoftResearch
MindJourney: Test-Time Scaling with World Models for Spatial Reasoning  @MicrosoftResearch
Uploaded August 2025 | Updated November 2026, 2 hours ago
The video introduces MindJourney, a framework that enhances Vision-Language Models (VLMs), which excel at interpreting single images but struggle to infer the underlying three-dimensional world. By allowing the VLM to “imagine” moving through the scene given a spatial reasoning question, the model proposes trajectories in a simulated imagination space. A world model then generates novel views along these paths, expanding the available observations from a single image. This richer 3D context enables the VLM to answer previously challenging questions with greater ease.

Publication: microsoft.com/en-us/research/publication/mindjourney-test-time-scaling-with-world-models-for-spatial-reasoning
MindJourney: Test-Time Scaling with World Models for Spatial ReasoningMagentic Marketplace: Testing societies of agents at scaleGenerative Gibbs sampling: Composing generative models with explicit physical contextEmail Spoofing with SMTP Smuggling: How the Shared Email Infrastructures Magnify this VulnerabilityIntroducing Quine: An AI research system designed for the complexity of biologyAI Testing and Evaluation: Learnings from genome editingTrailer: The Shape of Things to ComeDehumanizing machines: Making sense of AI systems that seem humanWhat AI gets wrong and what failure teaches usAbstracts: July 29, 2024Deep learning for DFTOptiMind: Teaching small language models to think like optimization experts
Microsoft Research |

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

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER