File size: 2,965 Bytes
20b00ad
 
 
 
 
 
 
 
 
 
 
 
7cbd5e5
20b00ad
 
 
 
337b2a7
 
 
20b00ad
337b2a7
 
b6247a7
337b2a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
20b00ad
 
337b2a7
20b00ad
 
 
 
 
 
 
 
 
 
337b2a7
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
---
library_name: transformers
tags:
- multimodal
- reasoning
- sft
- rl
datasets:
- LightChen2333/M3CoT
- ModalityDance/Omni-Bench
base_model:
- GAIR/Anole-7b-v0.1
pipeline_tag: any-to-any
---

# Omni-R1-Zero

[![Paper](https://img.shields.io/badge/Paper-arXiv-b31b1b?style=for-the-badge&logo=arxiv)](https://arxiv.org/abs/2601.09536)
[![Code](https://img.shields.io/badge/GitHub-Code-blue?style=for-the-badge&logo=github)](https://github.com/ModalityDance/Omni-R1)
[![Omni-Bench](https://img.shields.io/badge/Dataset-Omni--Bench-fcc21b?style=for-the-badge&logo=huggingface&logoColor=white)](https://huggingface.co/datasets/ModalityDance/Omni-Bench)

## Overview

**Omni-R1-Zero** is trained **without multimodal annotations**. It bootstraps **step-wise visualizations** from **text-only CoT seeds** (e.g., M3CoT), and then follows the same PeSFT+PeRPO recipe as Omni-R1 to learn interleaved multimodal reasoning.

## Usage

```python
import torch
from PIL import Image
from transformers import ChameleonProcessor, ChameleonForConditionalGeneration

# 1) Import & load
model_id = "ModalityDance/Omni-R1-Zero"  # or a local checkpoint path
processor = ChameleonProcessor.from_pretrained(model_id)
model = ChameleonForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

# 2) Prepare a single input
prompt = "You are a helpful assistant.\nUser: Which of these would appear shinier when polished? A. Metal spoon B. Wooden spoon\nThink with images first, the image reasoning process and answer are enclosed within <reserved12856> <reserved12857> and <reserved12866> <reserved12867> XML tags, respectively.\nAssistant:"

inputs = processor(
    prompt,
    padding=False,
    return_for_text_completion=True,
    return_tensors="pt",
).to(model.device)

# 3) Call the model
outputs = model.generate(
    **inputs,
    max_length=4096,
    do_sample=True,
    temperature=1.0,
    top_p=0.9,
    pad_token_id=1,
    multimodal_generation_mode="unrestricted",
)

# 4) Get results
text = processor.batch_decode(outputs, skip_special_tokens=False)[0]
print(text)
```

For full scripts (batch JSONL inference, interleaved decoding, and vLLM-based evaluation), please refer to the official GitHub repository:  
https://github.com/ModalityDance/Omni-R1

## License

This project is licensed under the **MIT License**.  
It also complies with the licenses of referenced third-party projects and dependencies, including the **Chameleon Research License**.

## Citation

```bibtex
@misc{cheng2026omnir1unifiedgenerativeparadigm,
      title={Omni-R1: Towards the Unified Generative Paradigm for Multimodal Reasoning}, 
      author={Dongjie Cheng and Yongqi Li and Zhixin Ma and Hongru Cai and Yupeng Hu and Wenjie Wang and Liqiang Nie and Wenjie Li},
      year={2026},
      eprint={2601.09536},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2601.09536}, 
}
```