cs.AIAug 25, 2026
SaveTask-Adaptive Rubrics for GUI Reward Modeling
Organizations: Zhejiang University · MiLM Plus, Xiaomi Inc.
Abstract
Recent studies on GUI agents have increasingly focused on outcome reward modeling, which assigns outcome rewards by judging whether an executed trajectory satisfies the success criteria implied by the user instruction. Existing GUI reward verifiers, however, often under-specify how these criteria should be constructed for each task instance. Whether using generic rubric structures or implicit model reasoning, their judging criteria are not sufficiently task-adaptive: they can transfer checks across tasks, overlook concrete constraints in the current instruction, or become overly strict by enforcing unstated requirements. To address this limitation, we propose AdaptRubric, a Coarse-to-Fine Rubrics Framework that constructs task-adaptive judging criteria through a category-level coarse stage and an instance-level fine stage. AdaptRubric performs category-level coarse rubric retrieval by routing the instruction to a GUI task family and retrieving reusable task-family criteria, then conducts instance-level fine rubric generation to surface compact cues for concrete values, scopes, and constraints in the current instruction. Across offline reward evaluation and online reinforcement learning optimization, AdaptRubric consistently outperforms prior reward agents, improving F1 by 3.6 points over the baseline average under a matched image budget and yielding a 4.23-point task-success gain.
Figures & tables
Figure 1: A motivating example for task-adaptive verification. The trajectory sets a related VS Code option to 50, but misses the requested line-length setting. While baselines verify only the surface value, AdaptRubric constructs coarse-to-fine criteria and correctly identifies the failure.
Figure 2: Overview of AdaptRubric . The left panel motivates task-adaptive verification with a concrete trajectory example and the resulting criterion gap. The middle panel illustrates coarse-to-fine rubric construction, where AdaptRubric retrieves a category-level coarse rubric from the rubric bank and augments it with compact instruction-specific cues. The right panel shows the final verification stage, where the fused criterion and selected trajectory context are passed to a VLM verifier for outcome reward prediction.
| Model | Ubuntu | Mobile | Windows | macOS | Web | Overall | ||||||||
| Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | Prec | Rec | F1 | |
| ZeroGUI | ||||||||||||||
| Qwen3-VL-4B-Instruct | 83.8 | 84.0 | 74.5 | 76.2 | 80.3 | 75.6 | 90.9 | 78.8 | 81.1 | 83.2 | 82.0 | 83.2 | 80.0 | 81.6 |
| Qwen3-VL-8B-Instruct | 84.6 | 85.0 | 83.0 | 84.5 | 79.3 | 74.4 | 94.8 | 88.2 | 78.4 | 80.6 | 83.3 | 84.1 | 81.9 | 83.0 |
| Qwen3-VL-32B-Instruct | 84.6 | 85.2 | 83.5 | 85.0 | 80.3 | 75.6 | 96.1 | 90.3 | 82.1 | 83.5 | 84.1 | 84.6 | 83.1 | 83.9 |
| Qwen3-VL-235B-A22B-Instruct | 86.9 | 87.3 | 84.0 | 85.8 | 84.5 | 80.9 | 96.1 | 90.3 | 87.4 | 88.6 | 86.7 | 87.2 | 85.9 | 86.5 |
Table 1: Offline reward discrimination results on OGRBench. Each method reports Accuracy (Acc) and F1-score (F1) for each platform; Overall columns include Acc, Precision (Prec), Recall (Rec), and F1. ZeroGUI uses the last-10 image budget. In each column, bold marks the best value across all method-backbone cells, and underline marks the second-best (computed separately for the Mean rows).
Figure 3: Cross-backbone radar comparison on OGRBench under the matched ten-screenshot budget. Each axis corresponds to one Qwen-family judge backbone, and each curve corresponds to one reward verifier.
| Backbone | Reward Agent | SR (%) | |
| MAI-UI-8B | – | 19.70 † | – |
| DigiRL | 23.08 | ||
| ZeroGUI | 22.22 | ||
| OS-Themis | 22.22 | ||
| AdaptRubric |
Table 2: Online reinforcement training results on MAI-UI-8B. We report the success rate (SR) after training with different reward agents. † denotes the result from Tang et al. (2026) .
Figure 4: Test-time scaling results on the heterogeneous AndroidWorld pool of 113 tasks.
| Method | ES@7 | BoN@8 | Acc. | F1 | FPR |
| OS-Themis | +2.15 | +7.97 | 75.75 | 72.92 | 10.81 |
| DigiRL | +7.46 | +7.97 | 80.80 | 81.90 | 22.71 |
| ZeroGUI | +10.11 | +12.39 | 86.55 | 87.16 | 15.38 |
| AdaptRubric | +11.88 | +13.28 | 88.14 | 88.41 | 11.17 |
Table 3: Heterogeneous-pool offline reward-guided test-time scaling. ES@7 (EarlyStop) and BoN@8 (BestOfN) report success-rate gains over Random in percentage points. Acc., F1, and FPR (false-positive rate) are computed over trajectory-level reward predictions.
| Coarse | Fine | Setting | Acc. | Acc. | F1 | F1 |
| Full (Coarse + Fine) | 86.9 | — | 86.6 | — | ||
| w/o Fine | 84.9 | 84.6 | ||||
| w/o Coarse | 85.2 | 84.6 | ||||
| w/o Both (image-only) | 81.5 | 80.7 |
Table 4: Ablation study of rubric construction on Qwen3.5-122B-A10B. Acc. and F1 report the absolute drop relative to the full model.
| Method | Quality | LLM Cost | Runtime | |||||
| Acc. | F1 | Calls | Calls/traj. | Tokens | Tok./traj. | Min. | s/traj. | |
| OS-Themis | 78.7 | 76.5 | 21,490 | 15.25 | 243.75M | 173.0K | 88.8 | 3.78 |
| ZeroGUI | 83.3 | 83.0 | 5,608 | 3.98 | 100.15M | 71.1K | 51.7 | 2.20 |
| DigiRL | 78.7 | 80.8 | 1,402 | 1.00 | 23.46M | 16.6K | 34.2 | 1.46 |
| AdaptRubric | 84.0 | 84.2 | 4,220 | 3.00 | 26.98M | 19.1K | 36.6 | 1.56 |
Table 5: Efficiency comparison on OGRBench using Qwen3-VL-8B-Instruct. All methods use the same ten-screenshot trajectory budget. Lowest cost in each column is in bold.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Family | Description | Key Dimensions |
| info_query | Answer from screen | Format; content; evidence |
| create_modify | Create or edit | Explicit properties; completion; consistency |
| delete_cleanup | Remove or clear | Scope; target; completeness |
| communication | Send to recipient | Source; recipient; channel; confirmation |
| transfer | Move or copy | Source; action; destination; fidelity |
| state_navigation | Toggle or navigate | Final state; toggle; persistence |
Table 6: GUI task taxonomy used for coarse rubric retrieval.
| Item | Value |
| Training framework | ClawGUI-RL |
| Environment | MobileWorld |
| Policy backbone | MAI-UI-8B |
| Optimization algorithm | GRPO |
| Training epochs | 2 |
| Number of GPUs | 8 |
Table 7: Key hyperparameters for online RL training.
| Setting | Acc | Prec | Rec | F1 |
| Last 2 screenshots | 79.1 | 84.9 | 70.7 | 76.8 |
| Last 10 screenshots | 85.0 | 84.9 | 84.9 | 84.9 |
Table 8: Effect of image budget on ZeroGUI. We report overall performance on OGRBench averaged across eight judge backbones.
| Model | Ubuntu | Mobile | Windows | macOS | Web | Overall | ||||||||
| Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | Prec | Rec | F1 | |
| DigiRL | ||||||||||||||
| Qwen3-VL-4B-Instruct | 70.7 | 71.0 | 78.7 | 81.3 | 76.5 | 72.5 | 87.0 | 70.6 | 76.3 | 79.6 | 74.3 | 74.1 | 74.1 | 74.1 |
| Qwen3-VL-8B-Instruct | 73.3 | 73.8 | 73.4 | 76.9 | 77.9 | 73.7 | 88.3 | 71.0 | 74.2 | 77.8 | 74.9 | 74.7 | 75.0 | 74.8 |
| Qwen3-VL-32B-Instruct | 77.9 | 78.1 | 80.9 | 82.5 | 79.8 | 74.9 | 90.9 | 75.9 | 81.1 | 82.9 | 79.7 | 81.1 | 77.1 | 79.1 |
| Qwen3-VL-235B-A22B-Instruct | 77.1 | 76.5 | 78.7 | 80.4 | 78.4 | 72.3 | 90.9 | 75.9 | 81.1 | 82.5 | 78.8 | 81.9 | 73.6 | 77.5 |
Table 9: Additional offline OGRBench results under each baseline’s original trajectory-context setting. Most baselines use only the final screenshot or a small final-state context, while ZeroGUI follows the prior two-screenshot setting.
| Baseline | Original trajectory context | Matched-budget instantiation | Verification style |
| ZeroGUI ( Yang et al., 2025 ) | Final-state selection; the OS-Themis evaluation protocol uses the last two screenshots. | Use the final ten screenshots and the same four-vote majority setting as the main offline comparison. | Terminal-state judgment with voting. |
| DigiRL ( Bai et al., 2024 ) | Autonomous evaluator prompt that judges task completion from visual evidence, typically centered on the observed screenshot/state. | Provide the same ten selected trajectory screenshots as visual evidence. | Autonomous trajectory evaluator. |
| DistRL ( Wang et al., 2025a ) | AUTO-EVALUATOR uses the task description with the last screenshot and recent action context. | Provide the same ten selected trajectory screenshots and the corresponding trajectory context. | Autonomous trajectory evaluator. |
| AndroidGen ( Lai et al., 2025 ) | StepCritic-style evaluation uses the task, action history, and final screen to decompose required conditions and check completion. | Provide the same ten selected trajectory screenshots together with the action history. | Condition-wise task checking. |
| WebRL ( Qi et al., 2025 ) | Outcome reward model uses the user intent, action history, and final web state. | Provide the same ten selected trajectory screenshots and trajectory history, replacing web-only state evidence with visual GUI evidence for cross-platform OGRBench. | Outcome reward modeling. |
| OS-Themis ( Li et al., 2026 ) | Milestone-based multi-agent critic that selects and verifies outcome-critical trajectory evidence. | Use the same ten-screenshot trajectory budget as AdaptRubric for offline comparison. | Structured multi-step evidence verification. |
Table 10: Baseline context budgets and matched-budget instantiations for the expanded OGRBench comparison. All matched-budget runs use at most ten trajectory screenshots to align with AdaptRubric .
| Model | Ubuntu | Mobile | Windows | macOS | Web | Overall | ||||||||
| Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | Prec | Rec | F1 | |
| DigiRL | ||||||||||||||
| Qwen3-VL-4B-Instruct | 77.5 | 81.2 | 76.1 | 80.5 | 77.0 | 77.4 | 88.3 | 76.9 | 72.1 | 77.6 | 77.1 | 70.7 | 92.0 | 80.0 |
| Qwen3-VL-8B-Instruct | 80.0 | 82.8 | 76.6 | 80.4 | 79.8 | 78.8 | 89.6 | 78.9 | 70.0 | 76.2 | 78.7 | 73.2 | 90.1 | 80.8 |
| Qwen3-VL-32B-Instruct | 85.2 | 86.8 | 82.4 | 85.2 | 82.6 | 80.6 | 94.8 | 88.2 | 79.5 | 82.7 | 84.2 | 79.6 | 91.6 | 85.2 |
| Qwen3-VL-235B-A22B-Instruct | 87.9 | 89.1 | 80.9 | 83.9 | 84.0 | 82.3 | 94.8 | 87.5 | 82.1 | 84.5 | 85.9 | 81.8 | 92.3 | 86.7 |
Table 11: Additional offline OGRBench results under the matched ten-screenshot trajectory budget. We include DigiRL, DistRL, AndroidGen, WebRL, ZeroGUI, OS-Themis, and AdaptRubric across multiple judge backbones.
| Pool | Trials/task | All-fail | All-pass | Mixed | Trial SR | Oracle@ | Headroom |
| 8B-only | 8 | 48 (42.5%) | 44 (38.9%) | 21 (18.6%) | 51.44% | 57.52% | +6.08 |
| 235B-only | 2 | 46 (40.7%) | 52 (46.0%) | 15 (13.3%) | 52.65% | 59.29% | +6.64 |
| 8B+235B | 10 | 32 (28.3%) | 29 (25.7%) | 52 (46.0%) | 51.68% | 71.68% | +20.00 |
Table 12: Candidate-pool composition for reward-guided test-time scaling. Mixed tasks contain both successful and failed trajectories, and are the only tasks where reward-guided selection can change the final task success. Headroom is Oracle@ minus Random@1.
Figure 5: Qualitative case study illustrating the importance of task-adaptive criteria. The task requires adding Hello, World! to the top of note_SiFbv.txt in Markor. The trajectory inserts the text, but places it below the original note content, yielding a failed task. Baseline verifiers focus on content presence and incorrectly judge the task as completed. AdaptRubric combines a coarse create/modify rubric with fine placement cues, enabling the verifier to check both the edited item and the instruction-specific requirement that no text appears above the inserted phrase.
More in this topic
Reward Modeling
Reward modeling learns scores or rankings that capture human goals and task success, guiding AI selection, reinforcement learning, and alignment.
See all 325 papers