Dream-VL & Dream-VLA: Open Vision-Language and Vision-Language-Action Models with Diffusion Language Model Backbone

TL;DR

This work investigates diffusion-based large language model backbones as a foundation for vision-language modeling (Dream-VL) and vision-language-action (Dream-VLA) in both general multimodal understanding and robotic control. Dream-VL leverages a diffusion LLM built on Dream-7B, trained on 12M open multimodal samples, and demonstrates competitive performance with autoregressive VLMs while excelling in planning tasks. Dream-VLA extends Dream-VL with large-scale robotic pretraining (Open-X Embodiment) and shows top-tier results on LIBERO, SimplerEnv, WidowX, and Google Robot benchmarks, with faster convergence during fine-tuning and robust action chunking without architectural changes. Together, these results highlight the potential of diffusion backbones for long-horizon planning and embodied AI, and the authors release both Dream-VL and Dream-VLA to spur community research.

Abstract

While autoregressive Large Vision-Language Models (VLMs) have achieved remarkable success, their sequential generation often limits their efficacy in complex visual planning and dynamic robotic control. In this work, we investigate the potential of constructing Vision-Language Models upon diffusion-based large language models (dLLMs) to overcome these limitations. We introduce Dream-VL, an open diffusion-based VLM (dVLM) that achieves state-of-the-art performance among previous dVLMs. Dream-VL is comparable to top-tier AR-based VLMs trained on open data on various benchmarks but exhibits superior potential when applied to visual planning tasks. Building upon Dream-VL, we introduce Dream-VLA, a dLLM-based Vision-Language-Action model (dVLA) developed through continuous pre-training on open robotic datasets. We demonstrate that the natively bidirectional nature of this diffusion backbone serves as a superior foundation for VLA tasks, inherently suited for action chunking and parallel generation, leading to significantly faster convergence in downstream fine-tuning. Dream-VLA achieves top-tier performance of 97.2% average success rate on LIBERO, 71.4% overall average on SimplerEnv-Bridge, and 60.5% overall average on SimplerEnv-Fractal, surpassing leading models such as $π_0$ and GR00T-N1. We also validate that dVLMs surpass AR baselines on downstream tasks across different training objectives. We release both Dream-VL and Dream-VLA to facilitate further research in the community.

Figures (8)

• Figure 1: Overview of the Dream family, with bi-directional masked diffusion modeling. Built on the diffusion language model Dream-7B, we introduce Dream-VL, a state-of-the-art diffusion VLM that demonstrates strong multimodal understanding, reasoning, and effective long-horizon planning. Building further on Dream-VL, we perform VLA pretraining to obtain the first pretrained diffusion-based VLA model (dVLA), which serves as a strong backbone for downstream VLA tasks.
• Figure 2: High-level and low-level action planning. Both forms of planning require multi-turn interaction with the environment, typically modeled as a Markov Decision Process. The task requires the model to perform long-horizon planning with high consistency. High-level planning operates in a symbolic or semantic action space, producing abstract commands. Low-level planning directly focuses on precise robot behavior, handling fine-grained, continuous control such as grasping and motion trajectories. Additional robotic training is required for bridging the vision-language model with executable robotic actions.
• Figure 3: Output formats and domains of the ViPlan benchmark.
• Figure 4: Performance comparison for VLMs on the ViPlan benchmark merler2025viplan.
• Figure 5: Comparison of model performance and speed with varying action chunk sizes.