Dream-VL & Dream-VLA

Building on the success of Dream 7B, we introduce Dream-VL and Dream-VLA, open VL and VLA models that fully unlock discrete diffusion’s advantages in long-horizon planning, bidirectional reasoning, and parallel action generation for multimodal tasks.

Key Results:

• Dream-VL: Achieves state-of-the-art performance among diffusion VLMs, comparable to top-tier AR VLMs trained on open data, with superior performance on visual planning tasks requiring long-horizon reasoning.
• Dream-VLA: Establishes top-tier performance with 97.2% average on LIBERO, 71.4% on SimplerEnv–Bridge, and 60.5% on SimplerEnv–Fractal, surpassing leading models including GR00T-N1 and OpenVLA-OFT. Consistently outperforms AR baselines across diverse finetuning objectives.

🚀 News

• 2025-12-23: Dream-VL & Dream-VLA models and blog released.

🧱 Repository Structure

The exact structure may evolve; please refer to the repo for up-to-date details.

Dream-VLX/
├── Dream-VL/          # Dream-VL training and evaluation (preparing)
├── Dream-VLA/         # Dream-VLA training and evaluation (preparing)
└── README.md          # This file

Citation

@article{ye2025dreamvla,
  title={Dream-VL & Dream-VLA: Open Vision-Language and Vision-Language-Action Models with Diffusion Language Model Backbone},
  author={Ye, Jiacheng and Gong, Shansan and Gao, Jiahui and Fan, Junming and Wu, Shuang and Bi, Wei and Bai, Haoli and Shang, Lifeng and Kong, Lingpeng},
  journal={arXiv preprint},
  year={2025}
}