Qwen2.5-VL-MullGRPO

• Repository: [https://github.com/arijitray1993/mull-tokens]
• Paper: [https://arxiv.org/abs/2512.10941]

How to Get Started with the Model

WORK IN PROGRESS: more details to be added soon!

It is highly recommended to install this version of transformers: https://github.com/arijitray1993/Mirage

git clone https://github.com/arijitray1993/Mirage
pip install -e ./transformers/.

Next, clone this repo: https://github.com/arijitray1993/mull-tokens.

We use a custom Qwen2.5 VL model. There is no change to the architecture, just some new tokens added.

% pip install qwen-vl-utils[decord]==0.0.8

import importlib
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor

Qwen2_5_VLForConditionalGeneration = importlib.import_module(
            'models.mmlatentdiscrete_qwen_vl'
        ).Qwen2_5_VLForConditionalGeneration

model = Qwen2_5_VLForConditionalGeneration.from_pretrained("array/Qwen2.5-VL-MullGRPO")
processor = AutoProcessor.from_pretrained(
    "array/Qwen2.5-VL-MullGRPO",
    trust_remote_code=True
)

Citation [optional]

@misc{ray2025mulltokensmodalityagnosticlatentthinking,
      title={Mull-Tokens: Modality-Agnostic Latent Thinking}, 
      author={Arijit Ray and Ahmed Abdelkader and Chengzhi Mao and Bryan A. Plummer and Kate Saenko and Ranjay Krishna and Leonidas Guibas and Wen-Sheng Chu},
      year={2025},
      eprint={2512.10941},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.10941}, 
}