🌍 World2VLM: Distilling World Model Imagination into VLMs for Dynamic Spatial Reasoning

📄 Paper • 💻 Code • 🤗 Dataset

---

✨ Overview

This repository provides a demo dataset for the paper:

World2VLM: Distilling World Model Imagination into VLMs for Dynamic Spatial Reasoning
Wanyue Zhang et al., 2026

🔍 Motivation
Vision-Language Models (VLMs) excel at static visual understanding but struggle with dynamic spatial reasoning, such as predicting how a scene changes under actions (e.g., moving forward, turning).

💡 Key Idea
We introduce World2VLM, a framework that uses world models as training-time teachers to distill spatial imagination into VLMs—enabling them to reason about future views and action consequences without external simulation at inference time.

📦 This repository contains:

• A compact demo dataset showcasing the data construction pipeline
• Representative trajectory-based supervision samples
• Examples of 8 dynamic spatial reasoning task types

⚠️ The full dataset will be released soon.

---

🧠 What is World2VLM?

World2VLM trains VLMs to mentally simulate the world by learning from world-model-generated transitions:

• Input: an image + an action (e.g., move forward)
• World model: generates the future view
• Output: structured supervision for reasoning

This enables two key capabilities:

• 🔁 Inverse reasoning: infer the action from image changes
• 🔮 Forward reasoning: predict what happens after an action

Unlike prior work, no world model is needed at inference time.

---

📂 Dataset Structure

data-demo/
├── README.md
├── SVC-RealScene-demo
│   ├── tasks_demo.jsonl
│   └── scenes/demo_scene/...
├── SVC-SimulatedScene-demo
│   ├── tasks_demo.jsonl
│   └── scenes/demo_scene/...
├── HY-WorldPlay-RealScene-demo
│   ├── tasks_demo.jsonl
│   └── scenes/demo_scene/...
└── HY-WorldPlay-SimulatedScene-demo
    ├── tasks_demo.jsonl
    └── scenes/demo_scene/...

🔍 Included Demo Subsets

We provide four compact subsets covering:

Teacher Model	Scene Type	Description
SVC	Real Scene	Camera-conditioned view synthesis
SVC	Simulated Scene	Synthetic environment transitions
HY-WorldPlay	Real Scene	Action-conditioned world dynamics
HY-WorldPlay	Simulated Scene	Long-horizon simulated trajectories

Each subset includes:

• 🎬 A trajectory bundle (images + metadata)
• 📝 A tasks_demo.jsonl file with structured supervision

---

🧾 Data Format

Each line in tasks_demo.jsonl represents one training example.

Common Fields

• task_type
  One of 8 spatial reasoning tasks: A1–A4, D1–D4

• messages
  A two-turn conversation:

  • User prompt
  • Target answer

• images
  Relative paths to referenced images

---

🧩 Task Suite (8 Types)

World2VLM defines a bidirectional task suite:

🔁 Motion-Centric (A-series)

Task	Description
A1	Motion distance estimation
A2	Motion orientation estimation
A3	Multi-step motion prediction
A4	Action-sequence verification

🎯 Object-Centric (D-series)

Task	Description
D1	Post-action bounding box prediction
D2	Post-action visibility detection
D3	Cross-view action inference
D4	Object consistency across views

💡 These tasks jointly enforce:

• Understanding camera motion
• Tracking object transformations
• Reasoning about viewpoint changes

---

⚙️ Data Construction Pipeline

The dataset is generated using world models as teachers:

1. 🖼️ Start from an anchor image
2. 🎮 Sample an egocentric action sequence
3. 🌐 Generate future views via world models
4. 🧠 Convert transitions into:
   • Forward tasks (predict outcomes)
   • Inverse tasks (recover actions)

This yields structured supervision of the form:

• P(action | before, after) (inverse)
• P(outcome | before, action) (forward)

---

🚀 Key Features

• 🧠 Spatial imagination distilled into VLMs
• 🔄 Bidirectional reasoning supervision
• 🧩 Multi-task structured dataset
• ⚡ No world model needed at inference
• 🌍 Supports both real and simulated scenes

---

📊 Why This Matters

World2VLM addresses a core limitation:

❌ VLMs fail at mental simulation
✅ World models can simulate—but are expensive

👉 Our solution:
Train VLMs to internalize world-model reasoning.

This leads to:

• Better dynamic spatial reasoning
• Lower inference cost
• Improved performance on spatial reasoning benchmarks

---

📝 Notes

• This repo contains demo-scale data only
• Full dataset (~100K samples) will be released soon
• Demo is intended for:
  • 🔍 Format inspection
  • 🧪 Pipeline understanding
  • 🧠 Task design exploration

---

📚 Citation

If you find this work useful, please cite:

@misc{zhang2026world2vlmdistillingworldmodel,
      title={World2VLM: Distilling World Model Imagination into VLMs for Dynamic Spatial Reasoning}, 
      author={Wanyue Zhang and Wenxiang Wu and Wang Xu and Jiaxin Luo and Helu Zhi and Yibin Huang and Shuo Ren and Zitao Liu and Jiajun Zhang},
      year={2026},
      eprint={2604.26934},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.26934}, 
}

---

🤝 Acknowledgements

We thank the community for advances in:

• 🌐 World Models (e.g., SVC(https://arxiv.org/abs/2503.14489), HY-WorldPlay(https://arxiv.org/abs/2412.03603))
• 🤖 Vision-Language Models
• 🧠 Spatial reasoning benchmarks

---

📬 Contact

For questions or collaborations, please open an issue or contact the authors via the paper.

---