| --- |
| language: |
| - en |
| license: apache-2.0 |
| pipeline_tag: image-text-to-text |
| base_model: |
| - k4ng/SCOPE-SFT-9B |
| tags: |
| - computer-use-agent |
| - gui-agent |
| - multimodal |
| - reinforcement-learning |
| - grpo |
| - safety |
| - osworld |
| - os-blind |
| --- |
| |
| # SCOPE-RL-9B |
|
|
| **SCOPE-RL-9B** is a computer-use agent (CUA) trained under the SCOPE (Safety and Capability Optimization for Policy Execution) framework to balance task-execution capability with safety-aware decision-making. The model is initialized from **SCOPE-SFT-9B** and further optimized through online reinforcement learning on verifiable capability tasks. |
|
|
| In our evaluation, SCOPE-RL-9B achieves a **54.17%** task success rate on OSWorld and a **64.30%** attack-avoidance rate on OS-BLIND, corresponding to a capability-safety harmonic mean of **58.80%**. Under our evaluation setting and among the models listed below, SCOPE-RL-9B achieves the best overall balance between capability and safety. |
|
|
| SCOPE-RL-9B is trained with programmatically verified capability tasks generated by SCOPE-Gen, using Safactory as the online reinforcement learning framework. Its initialization checkpoint, SCOPE-SFT-9B, was previously trained on capability demonstrations, safe-continuation trajectories, and explicit-refusal trajectories. |
|
|
| ## Links |
|
|
| - Paper: [Beyond Task Completion: Training Capable and Safe Computer-Use Agents]() |
| - Data generation code: [SCOPE-Gen](https://github.com/k4ngzy/SCOPE-Gen) |
| - RL training code: [Safactory](https://github.com/AI45Lab/SAfactory) |
| - Model collection: [SCOPE](https://huggingface.co/collections/k4ng/scope) |
|
|
| ## Quick Start |
|
|
| Install vLLM: |
|
|
| ```bash |
| pip install -U vllm |
| ``` |
|
|
| Launch an OpenAI-compatible inference server: |
|
|
| ```bash |
| vllm serve k4ng/SCOPE-RL-9B \ |
| --host 0.0.0.0 \ |
| --port 8000 \ |
| --tensor-parallel-size 1 \ |
| --data-parallel-size 2 \ |
| --trust-remote-code \ |
| --served-model-name scope-rl |
| ``` |
|
|
| ## Results |
|
|
| | Type | Model | H ↑ | OSWorld ↑ | OS-BLIND ↑ | |
| |---|---|---:|---:|---:| |
| | Closed-source | Claude 4.5 Sonnet | 37.78 | 62.90 | 27.00 | |
| | Closed-source | Qwen3.7-Plus | 9.36 | **73.33** | 5.00 | |
| | Open-source | EvoCUA-8B | 10.62 | 46.06 | 6.00 | |
| | Open-source | EvoCUA-32B | 4.42 | 56.73 | 2.30 | |
| | Open-source | OpenCUA-7B | 3.21 | 28.85 | 1.70 | |
| | Open-source | OpenCUA-32B | 1.94 | 34.79 | 1.00 | |
| | Open-source | OpenCUA-72B | 4.38 | 44.99 | 2.30 | |
| | Open-source | UI-TARS-1.5-7B | 8.46 | 27.52 | 5.00 | |
| | Open-source | ComputerRL | 21.77 | 48.90 | 14.00 | |
| | Open-source | Qwen3.5-9B | 8.93 | 41.80 | 5.00 | |
| | Open-source | Qwen3-VL-8B | 16.27 | 33.90 | 10.70 | |
| | SCOPE | SCOPE-Capability-Safety | 56.83 | 49.72 | **66.30** | |
| | SCOPE | **SCOPE-RL-9B** | **58.80** | **54.17** | 64.30 | |
|
|
| All values are percentages, and ↑ indicates that higher is better. \(H\) is the harmonic mean of the OSWorld task success rate and the OS-BLIND attack-avoidance rate. Compared with SCOPE-Capability-Safety, SCOPE-RL-9B improves OSWorld performance by 4.45 percentage points, while attack avoidance decreases by 2.00 points. The harmonic mean increases from 56.83% to 58.80%. |
|
|
| ## License |
|
|
| This model is subject to the license terms of its base model, Qwen3.5-9B. Licensing information for the code is available in the corresponding GitHub repositories. |
|
|
| ## Citation |
|
|
| If you use SCOPE-RL, SCOPE-SFT, SATraj-OS, or SCOPE-Gen, please cite: |
|
|
| ```bibtex |
| @misc{kang2026scope, |
| title = {Beyond Task Completion: Training Capable and Safe Computer-Use Agents}, |
| author = {Zeyu Kang and Zhenyun Yin and Yang Zhang and Shan He and Shanzhe Lei and Yanjiu Zhong and Xinquan Chen and Xuhong Wang}, |
| year = {2026} |
| } |
| ``` |
