The research team introduced BlueLM-GUI, a 35 billion parameter mobile GUI agent designed to address limitations in existing agent frameworks. The system’s architecture centers around a ‘flywheel’ approach, aiming to continuously improve performance through three core principles. First, ‘Every Sample Matters’ employs a dual-track pipeline, incorporating a Heterogeneous Triple-System Consensus evaluation and an Error Correction & Derivation Module to maximize the utility of training data. Second, ‘Every Rollout Is Real’ involves a three-stage training process: continual pre-training, supervised fine-tuning, and agentic reinforcement learning, all conducted on hundreds of real phones. This grounding in real-world production environments is intended to ensure the model’s capabilities translate directly to deployment. Finally, ‘Every Query Evolves’ utilizes a quota-driven benchmark methodology with three orthogonal axes to enable systematic model improvement.
The BlueLM-GUI model achieved a score of 87.4 on the MobileGUI-VBench benchmark, exceeding the performance of the best closed-source model by 5.1 points. Furthermore, it attained a score of 84.9 on the AndroidWorld benchmark, representing the highest score among open-source models and competitive with closed-source models. These results demonstrate the effectiveness of the real-device centric training approach. The system’s design focuses on robust and transferable mobile GUI capabilities.
The project involved 43 authors and utilized a continuous evaluation process. The system’s architecture is designed for iterative improvement, leveraging both real-device data and a dynamically updated benchmark. The team’s approach aims to reduce the mismatch between training and production environments, a common challenge in deploying AI models.
The BlueLM-GUI system is available as an open-weight model. The research team has provided code and data associated with the project through various tools, including CatalyzeX Code Finder for Papers and Influence Flower.
Source: https://arxiv.org/abs/2609.12394