cs.LGMar 25, 2026
SaveUI-Voyager: A Self-Evolving GUI Agent Learning via Failed Experience
Organizations: Tencent Hunyuan
Abstract
Autonomous mobile GUI agents have attracted increasing attention along with the advancement of Multimodal Large Language Models (MLLMs). However, existing methods still suffer from inefficient learning from failed trajectories and ambiguous credit assignment under sparse rewards for long-horizon GUI tasks. To that end, we propose UI-Voyager, a novel two-stage self-evolving mobile GUI agent. In the first stage, we employ Rejection Fine-Tuning (RFT), which enables the continuous co-evolution of data and models in a fully autonomous loop. The second stage introduces Group Relative Self-Distillation (GRSD), which identifies critical fork points in group rollouts and constructs dense step-level supervision from successful trajectories to correct failed ones. Extensive experiments on AndroidWorld show that our 4B model achieves an 81.0% Pass@1 success rate, outperforming numerous recent baselines and exceeding human-level performance. Ablation and case studies further verify the effectiveness of GRSD. Our method represents a significant leap toward efficient, self-evolving, and high-performance mobile GUI automation without expensive manual data annotation.
Figures & tables
Figure 1: Performance comparison of various GUI agents on AndroidWorld. Our UI-Voyager (4B) achieves an 81.0% Pass@1 success rate, outperforming larger models and exceeding reported human-level performance.
Figure 2: The whole pipeline of training UI-Voyager for mobile GUI tasks. It consists of two iterative stages: (1) Rejection Fine-Tuning (RFT), where a base policy generates multiple trajectories that are filtered by a rule-based verifier to collect high-quality samples for supervised fine-tuning; (2) Group Relative Self-Distillation (GRSD), which identifies “fork points” between successful and failed trajectory groups using SSIM matching and corrects erroneous actions to further refine the policy through mixed-data training.
| Code Actions | Descriptions |
|---|---|
| click(x,y) | Clicking at coordinates (x,y) |
| long_press(x,y) | Long-pressing at coordinates (x,y) |
| swipe(x,y,x’,y’) | Swiping from (x,y) to (x’,y’) |
| open_app(app_name) | Opening an app by name |
| input_text(text) | Typing input text |
| keyboard_enter() | Pressing the Enter key on the keyboard |
Table 1: Predefined action space of AndroidWorld ( Rawles et al., 2025 ) .
Figure 3: Illustration of the fork point detection strategy. Given a successful trajectory and a failed trajectory for the same task, the fork point detection mechanism identifies steps in the failed trajectory where the screen state matches that of a successful step ( ) but the subsequent action leads to divergence ( ), indicating that the action taken in the failed trajectory deviates from the successful one. See Sec. 3.3.1 for details.
| Model | #Params | Success Rate |
|---|---|---|
| Baselines | ||
| Qwen3-VL-2B ( Bai et al., 2025a ) | 2B | 36.4 |
| MAI-UI-2B ( Zhou et al., 2025b ) | 2B | 49.1 |
| ScaleCUA-3B ( Liu et al., 2025 ) | 3B | 23.7 |
| Ferret-UI Lite-3B ( Yang et al., 2025c ) | 3B | 28.0 |
| Qwen3-VL-4B ( Bai et al., 2025a ) | 4B | 45.3 |
Table 2: Performance comparison on AndroidWorld Benchmark. Best results are in bold , and second-best results are underlined . UI-Voyager achieves an 81.0% success rate, surpassing all baseline methods and the reported human-level performance of 80.0%. Notably, our model achieves superior results with only 4B parameters, demonstrating strong efficiency compared to much larger models. To ensure reproducibility, we report the average success rate over 64 random seeds, whereas baseline results are taken from prior papers.
Figure 4: RFT significantly boosts agent performance. Left: Pass@K performance across four iterative rounds of RFT. The results show consistent improvement in both Pass@1 and Pass@k as the self-evolution progresses. We select the checkpoint from the third RFT round (Pass@1=73.2%) for subsequent training. Right: Training curves of GRPO and PPO initialized from Qwen3-VL-4B-Instruct. The results show that directly deploying RL algorithms from Qwen3-VL-4B-Instruct model yields marginal gains and exhibits high sample inefficiency.
Figure 5: Example of fork point detection on BrowserMaze task. Both failed and successful trajectories share the same screen state at Step 12 (fork point). The failed trajectory takes an invalid “Right” action (blocked by a wall), while the successful trajectory takes the correct “Down” action. The fork point detection mechanism identifies this divergence and uses the correct action from the successful trajectory to supervise the failed one at this critical step.
Figure 6: Example of fork point detection on SystemBluetoothTurnOff task. The fork point occurs at Step 0 (initial state), where both trajectories start from the same home screen. The failed trajectory attempts to open the settings app with an incorrect upward swipe, while the successful trajectory uses a downward swipe to open the notification shade and access quick settings. Fork point detection identifies this initial divergence, providing corrective supervision at the very first step.
Figure 7: Illustration of the self-corrective samples. Both failed and successful trajectories share the same screen state at Step 12 (fork point). The failed trajectory incorrectly takes the “Right” action with flawed reasoning (“I need to continue moving ‘X’ right and then down”), while the successful trajectory correctly takes the “Down” action with proper reasoning (“The next logical step is to move ‘X’ down towards the bottom-right corner”). Fork point detection identifies this divergence, enabling the construction of self-corrective samples that transform failed trajectories into high-quality supervised data.
Figure 8: Training performance comparison of GRSD, GRPO, and PPO. All methods start from the same RFT model with 73.2% success rate. GRSD successfully boosts the agent’s Pass@1 performance to 81% (Left), while GRPO and PPO show slower progress and plateau around 76% (Right). The results demonstrate that GRSD’s fork point detection and self-correction mechanisms enable more effective learning compared to standard RL baselines.
Figure 9: Performance comparison of GRSD, GRPO, and PPO on ten representative low-success-rate tasks. GRSD consistently achieves the highest success rate across all tasks, significantly outperforming both PPO and GRPO. In contrast, PPO and GRPO struggle to make substantial gains due to the scarcity of successful samples and the lack of effective credit assignment mechanisms. These results demonstrate that GRSD enables efficient learning from failed trajectories even in sparse-reward environments.
More in this topic
Mobile GUI Agents
Mobile GUI agents use visual and structured interface information to carry out tasks across smartphone apps.
See all 81 papers