---
license: mit
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- gui-agent
- rlvr
- computer-use
---

# BEPA-7B-S2

This repository contains the weights for **BEPA-7B-S2**, an end-to-end screenshot-to-action policy for GUI agents. The model was introduced in the paper [From Off-Policy to On-Policy: Enhancing GUI Agents via Bi-level Expert-to-Policy Assimilation](https://huggingface.co/papers/2601.05787).

## Introduction

**BEPA** (Bi-Level Expert-to-Policy Assimilation) is a framework designed to enhance Vision-Language Models acting as computer-use agents (CUAs). It addresses the challenges of using static expert trajectories in reinforcement learning from verifiable rewards (RLVR) by turning them into policy-aligned guidance.

BEPA operates in two complementary stages:
- **LEVEL-1 (Self-Rolled Execution):** Transforms alien expert traces into policy-compatible trajectories by abstracting them into natural-language plans and letting the base policy execute them.
- **LEVEL-2 (Self-Aligned Assimilation):** Dynamically maintains a per-task cache that injects guided trajectories into training updates when on-policy failures occur.

On the OSWorld-Verified benchmark, BEPA improves the success rate of UITARS1.5-7B from 22.87% to **32.13%**, establishing it as a top-performing open-source end-to-end model.

## Resources
- **Paper:** [https://huggingface.co/papers/2601.05787](https://huggingface.co/papers/2601.05787)
- **Project Page:** [https://leon-gittech.github.io/Verl_GUI/](https://leon-gittech.github.io/Verl_GUI/)
- **GitHub Repository:** [https://github.com/LEON-gittech/Verl_GUI](https://github.com/LEON-gittech/Verl_GUI)

## Main Results

| Method | Overall Success (%) |
|--------|-------------|
| UITARS1.5-7B | 22.87 |
| GRPO | 23.60 |
| **BEPA (ours)** | **32.13** |

## Citation

```bibtex
@misc{wang2026offpolicyonpolicyenhancinggui,
      title={From Off-Policy to On-Policy: Enhancing GUI Agents via Bi-level Expert-to-Policy Assimilation}, 
      author={Zezhou Wang and Ziyun Zhang and Xiaoyi Zhang and Zhuzhong Qian and Yan Lu},
      year={2026},
      eprint={2601.05787},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2601.05787}, 
}
```