README.md

---
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}
}
```