BlueLM-GUI: A Real-Device Flywheel for Self-Improving Mobile GUI Agents

Summary

The BlueLM-GUI technical report presents a 35B-A3B mobile GUI agent and a training process designed around real production devices. It targets three deployment gaps: the distribution mismatch caused by sandbox training, the loss of useful information from expensive real-device failures, and benchmark saturation that makes iteration harder to guide. The system follows three principles. “Every Sample Matters” combines Heterogeneous Triple-System Consensus evaluation with an Error Correction & Derivation Module to turn trajectories into supervision, including otherwise unsuccessful rollouts. “Every Rollout Is Real” uses continual pre-training, supervised fine-tuning, and agentic reinforcement learning across hundreds of real phones, aiming to make learned behavior transfer directly to deployment. “Every Query Evolves” uses quota-driven benchmarks with three orthogonal axes so results can be attributed more precisely and the benchmark can be upgraded as capabilities improve. BlueLM-GUI scores 87.4 on MobileGUI-VBench, 5.1 points above the best closed-source model reported in the paper. It also scores 84.9 on AndroidWorld, which the report describes as the best result among open-source models and competitive with closed-source systems. The results support the report’s claim that combining real-device training with iterative data and benchmark improvement produces strong and transferable mobile GUI capability.