---
license: cc-by-4.0

language:
  - en
tags:
  - visual-reasoning
  - unified-model
  - reinforcement-learning
  - emu3.5
  - multimodal
  - next-token-prediction
  - grpo
pipeline_tag: image-text-to-text
library_name: transformers
base_model:
  - BAAI/Emu3.5
datasets:
  - maverickrzw/VR-X-SFT-RL
---

# UniVR: Thinking in Visual Space for Unified Visual Reasoning

<p align="center">
  <img src="asset/Fig1_v1.png" alt="UniVR Overview" width="95%">
</p>

<p align="center">
  <a href="https://maverickren.github.io/UniVR.github.io/">🌐 Project Page</a> &nbsp;|&nbsp;
  <a href="#">📄 Paper</a> &nbsp;|&nbsp;
  <a href="https://github.com/MaverickRen/UniVR">💻 Code</a> &nbsp;|&nbsp;
  <a href="https://huggingface.co/datasets/maverickrzw/VR-X-SFT-RL">📦 VR-X Dataset</a>
</p>

---

## Model Summary

**UniVR** is the first framework that simultaneously learns complex reasoning, fine-grained physical dynamics, and long-term planning from pure visual demonstrations — without relying on dense image-text pairs or task-specific heuristics.

Built on [Emu3.5](https://huggingface.co/BAAI/Emu3.5) (34B), UniVR uses a unified next-token prediction objective to directly generate visual reasoning traces given an image and instruction. Training employs a two-stage pipeline: supervised cold initialization on the VR-X dataset, followed by **VR-GRPO** reinforcement learning with complementary global and step-focal rewards.

| Feature | Detail |
|---|---|
| **Architecture** | Emu3.5 34B (VQ-VAE unified generative model) |
| **Training** | SFT (310k samples) → VR-GRPO RL (3k samples) |
| **Visual Thinking** | Native visual-space reasoning, no intermediate text chain |
| **Benchmark** | VR-X: 16 sources, 6 task categories, 1.8k evaluation samples |

---

## Available Checkpoints

| Model | Description | Link |
|---|---|---|
| **UniVR-34B-Planning** | Optimized for long-horizon planning tasks (robotic manipulation, tool use, multi-step control) | [maverickrzw/UniVR-34B-Planning](https://huggingface.co/maverickrzw/UniVR-34B-Planning) |
| **UniVR-34B-General** | Full UniVR recipe with interleaved image-text data; suitable for general visual reasoning | [maverickrzw/UniVR-34B-General](https://huggingface.co/maverickrzw/UniVR-34B-General) |

---

## Key Results

### VR-X Benchmark

UniVR achieves up to **25% improvement** over the Emu3.5 baseline and approaches Gemini 3 Pro + Nano Banana 2 with only 34B parameters.

| Method | Visual Thinking | Guidance | Robot | Editing | Spatial | Puzzle | Search | Overall↑ |
|---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
| Gemini-3-pro + Nano Banana 2 | ✗ | 66.2 | 67.1 | 63.7 | 55.1 | 65.5 | 79.0 | **66.1** |
| GPT-5 + GPT-image-1.5 | ✗ | 68.2 | 64.1 | 58.0 | 49.3 | 64.0 | 77.4 | 63.5 |
| Emu3.5 34B | ✗ | 38.6 | 42.8 | 32.7 | 35.3 | 43.4 | 46.2 | 39.8 |
| **UniVR 34B** | **✓** | **59.5** | **68.0** | **48.5** | **46.5** | **62.2** | **64.3** | **58.2** |
| *Δ v.s. Emu3.5* | | *↑20.9* | *↑25.2* | *↑15.8* | *↑11.2* | *↑18.8* | *↑18.1* | *↑18.4* |

### Multimodal Understanding

Enhanced visual reasoning also boosts standard multimodal benchmarks — no degradation of the base model's capabilities.

| Method | MMMU | MME(P) | MME(C) | MMBench | MathVista | MM-Vet |
|---|:---:|:---:|:---:|:---:|:---:|:---:|
| Emu 3.5 | 0.292 | 781.1 | 324.6 | 0.183 | 41.7 | 28.0 |
| **UniVR** | **0.337** | **799.3** | **338.5** | **0.198** | **44.0** | **35.6** |
| *Δ v.s. Emu3.5* | *↑0.045* | *↑18.2* | *↑13.9* | *↑0.015* | *↑2.3* | *↑7.6* |

---

## Quick Start

### Installation

```bash
git clone https://github.com/MaverickRen/UniVR.git
cd UniVR
bash install.sh
```

### Inference

```bash
cd UniVR_SFT

# Download checkpoint
huggingface-cli download maverickrzw/UniVR-34B-Planning --local-dir weights/UniVR-34B-Planning

# Download VisionTokenizer
huggingface-cli download BAAI/Emu3.5-VisionTokenizer --local-dir weights/Emu3.5-VisionTokenizer

# Run inference
bash scripts/inference.sh
```

Configure `configs/config.py` to set model paths and prompts:

```python
{
    "prompt": "Tie the red rope around the white gift box. Finish this task in 3 steps.",
    "reference_image": "path/to/first_frame.jpg",
}
```

### Training

**SFT (Cold Initialization)**:
```bash
cd UniVR_SFT
# LoRA (2 nodes × 8 GPUs)
bash scripts/train_sft_lora.sh
# Full parameter (4 nodes × 8 GPUs)
bash scripts/train_sft_full.sh
```

**RL (VR-GRPO)**:
```bash
cd UniVR_RL
bash examples/emu3_grpo_lora.sh
```

---

## Method: VR-GRPO

UniVR proposes **VR-GRPO** (Visual Reasoning GRPO), a reinforcement learning paradigm that combines:

- **Global Reward (R_g)**: A VLM evaluator assesses overall task completion and visual quality via pairwise comparison.
- **Step-Focal Reward (R_s)**: Identifies the most error-prone sub-steps by computing inter-trajectory CLIP-feature variance across rollout samples, then applies fine-grained VLM evaluation on critical windows.
- **Combined Reward**: `R_reason = R_g − λ|R_g − R_s|`, enforcing both terminal correctness and procedural integrity.

This design prevents reward hacking in long-horizon tasks where global-only rewards overlook intermediate physical violations and logical gaps.

---

## Sample Outputs

<table>
<tr>
<td align="center"><b>Tie a Knot</b></td>
<td align="center"><b>Hang Clothes</b></td>
<td align="center"><b>Draw</b></td>
</tr>
<tr>
<td><img src="asset/tie_rope_02.jpg" width="250"/></td>
<td><img src="asset/hang_clothes_03.jpg" width="250"/></td>
<td><img src="asset/Draw.png" width="250"/></td>
</tr>
</table>

---

## Training Data

UniVR is trained on **VR-X**, a large-scale benchmark curated from 1.5M raw samples across 16 diverse sources:

| Category | Sources | Examples |
|---|---|---|
| Visual Guidance | EgoDex, Action100M, Epic-Kitchen, VideoCraftBench | Cooking, handcrafting, daily activities |
| Robot Manipulation | AgiBot, Droid, Bridge, ZebraCoT-Robot | Robotic grasping, tool use, multi-step control |
| Editing | ZebraCoT-Multiobject | Object manipulation, scene editing |
| Spatial Perception | ThinkMorph-Navigation, ZebraCoT-Embodiment | Navigation, spatial reasoning |
| Visual Search | VisualCoT, ThinkMorph-Search | Object localization, attention |
| Puzzle & Game | VRBench, Zebra-Jigsaw, ThinkMorph-VisPuzzle | Mazes, jigsaw, visual puzzles |

Download: [maverickrzw/VR-X-SFT-RL](https://huggingface.co/datasets/maverickrzw/VR-X-SFT-RL)

---

## Citation

```bibtex
@article{ren2026univr,
  title={UniVR: Thinking in Visual Space for Unified Visual Reasoning},
  author={Zhongwei Ren and Yunchao Wei and Zhao Yao and Guixun Luo and Yao Zhao and Weibo Gong and Xiao Liu and Anran Wang and Xiangtai Li and Xiaojie Jin},
  year={2026},
}
```

## License

This project is released under the CC BY 4.0 License.

## Acknowledgements

UniVR is built upon [Emu3.5](https://github.com/baaivision/Emu3) and [verl](https://github.com/volcengine/verl). We thank the authors for their excellent open-source contributions.