Datasets:

Modalities:
Image
Size:
< 1K
ArXiv:
Libraries:
Datasets
License:
World2VLM / README.md
WanyueZhang's picture
Update README.md
31842c6 verified
---
license: apache-2.0
---
# ๐ŸŒ World2VLM: Distilling World Model Imagination into VLMs for Dynamic Spatial Reasoning
<p align="center">
<a href="https://arxiv.org/abs/2604.26934">๐Ÿ“„ Paper</a> โ€ข
<a href="https://github.com/WanyueZhang-ai/World2VLM">๐Ÿ’ป Code</a> โ€ข
<a href="https://huggingface.co/datasets/WanyueZhang/World2VLM">๐Ÿค— Dataset</a>
</p>
---
## โœจ 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*
๐Ÿ” **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).
๐Ÿ’ก **Key Idea**
We introduce **World2VLM**, a framework that uses **world models as training-time teachers** to distill *spatial imagination* into VLMsโ€”enabling them to reason about **future views and action consequences without external simulation at inference time**.
๐Ÿ“ฆ **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**
โš ๏ธ The **full dataset will be released soon**.
---
## ๐Ÿง  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:
- ๐Ÿ” **Inverse reasoning**: infer the action from image changes
- ๐Ÿ”ฎ **Forward reasoning**: predict what happens after an action
Unlike prior work, **no world model is needed at inference time**.
---
## ๐Ÿ“‚ Dataset Structure
```bash
data-demo/
โ”œโ”€โ”€ README.md
โ”œโ”€โ”€ SVC-RealScene-demo
โ”‚ โ”œโ”€โ”€ tasks_demo.jsonl
โ”‚ โ””โ”€โ”€ scenes/demo_scene/...
โ”œโ”€โ”€ SVC-SimulatedScene-demo
โ”‚ โ”œโ”€โ”€ tasks_demo.jsonl
โ”‚ โ””โ”€โ”€ scenes/demo_scene/...
โ”œโ”€โ”€ HY-WorldPlay-RealScene-demo
โ”‚ โ”œโ”€โ”€ tasks_demo.jsonl
โ”‚ โ””โ”€โ”€ scenes/demo_scene/...
โ””โ”€โ”€ HY-WorldPlay-SimulatedScene-demo
โ”œโ”€โ”€ tasks_demo.jsonl
โ””โ”€โ”€ scenes/demo_scene/...
```
## ๐Ÿ” 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:
- ๐ŸŽฌ A **trajectory bundle** (images + metadata)
- ๐Ÿ“ A **`tasks_demo.jsonl`** file with structured supervision
---
## ๐Ÿงพ Data Format
Each line in `tasks_demo.jsonl` represents one training example.
### Common Fields
- `task_type`
One of 8 spatial reasoning tasks: `A1โ€“A4`, `D1โ€“D4`
- `messages`
A two-turn conversation:
- User prompt
- Target answer
- `images`
Relative paths to referenced images
---
## ๐Ÿงฉ Task Suite (8 Types)
World2VLM defines a **bidirectional task suite**:
### ๐Ÿ” Motion-Centric (A-series)
| Task | Description |
|------|-------------|
| A1 | Motion distance estimation |
| A2 | Motion orientation estimation |
| A3 | Multi-step motion prediction |
| A4 | Action-sequence verification |
### ๐ŸŽฏ 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 |
๐Ÿ’ก These tasks jointly enforce:
- Understanding **camera motion**
- Tracking **object transformations**
- Reasoning about **viewpoint changes**
---
## โš™๏ธ Data Construction Pipeline
The dataset is generated using **world models as teachers**:
1. ๐Ÿ–ผ๏ธ Start from an **anchor image**
2. ๐ŸŽฎ Sample an **egocentric action sequence**
3. ๐ŸŒ Generate **future views** via world models
4. ๐Ÿง  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)
---
## ๐Ÿš€ Key Features
- ๐Ÿง  **Spatial imagination distilled into VLMs**
- ๐Ÿ”„ **Bidirectional reasoning supervision**
- ๐Ÿงฉ **Multi-task structured dataset**
- โšก **No world model needed at inference**
- ๐ŸŒ Supports both **real and simulated scenes**
---
## ๐Ÿ“Š Why This Matters
World2VLM addresses a core limitation:
> โŒ VLMs fail at mental simulation
> โœ… World models can simulateโ€”but are expensive
๐Ÿ‘‰ **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
---
## ๐Ÿ“ Notes
- This repo contains **demo-scale data only**
- Full dataset (~100K samples) will be released soon
- Demo is intended for:
- ๐Ÿ” Format inspection
- ๐Ÿงช Pipeline understanding
- ๐Ÿง  Task design exploration
---
## ๐Ÿ“š 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},
}
```
---
## ๐Ÿค Acknowledgements
We thank the community for advances in:
* ๐ŸŒ World Models (e.g., SVC(https://arxiv.org/abs/2503.14489), HY-WorldPlay(https://arxiv.org/abs/2412.03603))
* ๐Ÿค– Vision-Language Models
* ๐Ÿง  Spatial reasoning benchmarks
---
## ๐Ÿ“ฌ Contact
For questions or collaborations, please open an issue or contact the authors via the paper.
---