| --- |
| license: apache-2.0 |
| --- |
| # ๐ World2VLM: Distilling World Model Imagination into VLMs for Dynamic Spatial Reasoning |
|
|
| <p align="center"> |
| <a href="https://arxiv.org/abs/2604.26934">๐ Paper</a> โข |
| <a href="https://github.com/WanyueZhang-ai/World2VLM">๐ป Code</a> โข |
| <a href="https://huggingface.co/datasets/WanyueZhang/World2VLM">๐ค Dataset</a> |
| </p> |
|
|
| --- |
|
|
| ## โจ 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* |
|
|
| ๐ **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). |
|
|
| ๐ก **Key Idea** |
| We introduce **World2VLM**, a framework that uses **world models as training-time teachers** to distill *spatial imagination* into VLMsโenabling them to reason about **future views and action consequences without external simulation at inference time**. |
|
|
| ๐ฆ **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** |
|
|
| โ ๏ธ The **full dataset will be released soon**. |
|
|
| --- |
|
|
| ## ๐ง 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: |
| - ๐ **Inverse reasoning**: infer the action from image changes |
| - ๐ฎ **Forward reasoning**: predict what happens after an action |
|
|
| Unlike prior work, **no world model is needed at inference time**. |
|
|
| --- |
|
|
| ## ๐ Dataset Structure |
|
|
| ```bash |
| data-demo/ |
| โโโ README.md |
| โโโ SVC-RealScene-demo |
| โ โโโ tasks_demo.jsonl |
| โ โโโ scenes/demo_scene/... |
| โโโ SVC-SimulatedScene-demo |
| โ โโโ tasks_demo.jsonl |
| โ โโโ scenes/demo_scene/... |
| โโโ HY-WorldPlay-RealScene-demo |
| โ โโโ tasks_demo.jsonl |
| โ โโโ scenes/demo_scene/... |
| โโโ HY-WorldPlay-SimulatedScene-demo |
| โโโ tasks_demo.jsonl |
| โโโ scenes/demo_scene/... |
| ``` |
|
|
| ## ๐ 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: |
| - ๐ฌ A **trajectory bundle** (images + metadata) |
| - ๐ A **`tasks_demo.jsonl`** file with structured supervision |
| |
| --- |
| |
| ## ๐งพ Data Format |
| |
| Each line in `tasks_demo.jsonl` represents one training example. |
| |
| ### Common Fields |
| |
| - `task_type` |
| One of 8 spatial reasoning tasks: `A1โA4`, `D1โD4` |
| |
| - `messages` |
| A two-turn conversation: |
| - User prompt |
| - Target answer |
| |
| - `images` |
| Relative paths to referenced images |
| |
| --- |
| |
| ## ๐งฉ Task Suite (8 Types) |
| |
| World2VLM defines a **bidirectional task suite**: |
| |
| ### ๐ Motion-Centric (A-series) |
| |
| | Task | Description | |
| |------|-------------| |
| | A1 | Motion distance estimation | |
| | A2 | Motion orientation estimation | |
| | A3 | Multi-step motion prediction | |
| | A4 | Action-sequence verification | |
| |
| ### ๐ฏ 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 | |
| |
| ๐ก These tasks jointly enforce: |
| - Understanding **camera motion** |
| - Tracking **object transformations** |
| - Reasoning about **viewpoint changes** |
| |
| --- |
| |
| ## โ๏ธ Data Construction Pipeline |
| |
| The dataset is generated using **world models as teachers**: |
| |
| 1. ๐ผ๏ธ Start from an **anchor image** |
| 2. ๐ฎ Sample an **egocentric action sequence** |
| 3. ๐ Generate **future views** via world models |
| 4. ๐ง 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) |
| |
| --- |
| |
| ## ๐ Key Features |
| |
| - ๐ง **Spatial imagination distilled into VLMs** |
| - ๐ **Bidirectional reasoning supervision** |
| - ๐งฉ **Multi-task structured dataset** |
| - โก **No world model needed at inference** |
| - ๐ Supports both **real and simulated scenes** |
| |
| --- |
| |
| ## ๐ Why This Matters |
| |
| World2VLM addresses a core limitation: |
| |
| > โ VLMs fail at mental simulation |
| > โ
World models can simulateโbut are expensive |
| |
| ๐ **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 |
| |
| --- |
| |
| ## ๐ Notes |
| |
| - This repo contains **demo-scale data only** |
| - Full dataset (~100K samples) will be released soon |
| - Demo is intended for: |
| - ๐ Format inspection |
| - ๐งช Pipeline understanding |
| - ๐ง Task design exploration |
| |
| --- |
| ## ๐ 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}, |
| } |
| ``` |
| |
| --- |
| ## ๐ค Acknowledgements |
| |
| We thank the community for advances in: |
| |
| * ๐ World Models (e.g., SVC(https://arxiv.org/abs/2503.14489), HY-WorldPlay(https://arxiv.org/abs/2412.03603)) |
| * ๐ค Vision-Language Models |
| * ๐ง Spatial reasoning benchmarks |
| |
| --- |
| ## ๐ฌ Contact |
| |
| For questions or collaborations, please open an issue or contact the authors via the paper. |
| |
| --- |