source: "raw/articles/ui-voyager-a-self-evolving-gui-agent-learning-via-failed-experience.md"

Summary: UI-Voyager: A Self-Evolving GUI Agent Learning via Failed Experience

TL;DR: A mobile GUI agent that achieves 81.0% success rate on AndroidWorld by learning from failed trajectories through fork point detection and self-distillation, outperforming larger models and human performance.

Key Points

• Performance: 4B model achieves 81.0% Pass@1 success rate on AndroidWorld, exceeding human performance (80.0%) and all baseline methods including much larger models (235B parameters)
• Two-stage training framework:
  • Stage 1: Rejection Fine-Tuning (RFT) for automatic data-model co-evolution
  • Stage 2: Group Relative Self-Distillation (GRSD) for learning from failed trajectories
• Fork point detection: Uses SSIM-based matching to identify critical decision points where successful and failed trajectories diverge
• Credit assignment solution: Addresses sparse reward problem in long-horizon GUI tasks by providing dense step-level supervision
• Self-corrective learning: Transforms failed trajectories into high-quality training data without manual annotation
• Evaluation: Tested on 116 diverse AndroidWorld tasks across real-world mobile applications
• Comparison: GRSD significantly outperforms standard RL methods (GRPO, PPO) which plateau around 76%

Concepts Covered

• Multimodal Large Language Models — backbone architecture using Qwen3-VL-4B-Instruct
• Reinforcement Learning — addresses credit assignment challenges in sparse reward environments
• Group Relative Policy Optimization — baseline RL method that GRSD outperforms
• AndroidWorld Benchmark — evaluation environment with 116 diverse mobile GUI tasks
• Rejection Sampling — filtering mechanism for high-quality trajectory collection
• Self-Distillation — knowledge transfer from successful to failed trajectories
• SSIM Image Matching — computer vision technique for identifying equivalent screen states
• Mobile GUI Automation — target application domain for autonomous phone operation
• Credit Assignment Problem — fundamental RL challenge addressed by fork point detection
• Self-Evolving Training — iterative improvement without manual data annotation

Images and Figures

• Figure 1 (raw/articles/2603.24533v1/x1.png): Performance comparison showing UI-Voyager achieving 81.0% vs other models
• Figure 2 (raw/articles/2603.24533v1/x2.png): Training pipeline overview showing RFT and GRSD stages
• Figure 3 (raw/articles/2603.24533v1/x3.png): Fork point detection illustration with successful/failed trajectory comparison
• Figure 4 (raw/articles/2603.24533v1/x4.png): RFT performance improvements and RL baseline comparisons
• Figure 5 (raw/articles/2603.24533v1/x5.png): BrowserMaze task example showing fork point at step 12
• Figure 6 (raw/articles/2603.24533v1/x6.png): SystemBluetoothTurnOff task example with fork point at step 0
• Figure 7 (raw/articles/2603.24533v1/x7.png): Self-corrective sample construction process
• Figure 8 (raw/articles/2603.24533v1/x8.png): GRSD vs GRPO/PPO training performance comparison
• Figure 9 (raw/articles/2603.24533v1/x9.png): Performance on low-success-rate tasks

Related Concepts

• Vision-Language Models
• Mobile App Automation
• GUI Understanding
• Computer Vision for UI
• Long-Horizon Planning
• Sparse Reward Learning
• Self-Supervised Learning
• Trajectory Optimization
• Interactive Environment Training
• Multi-Turn Dialogue Systems