README.md

---
base_model: Qwen/Qwen3-VL-8B-Instruct
datasets:
- lgy0404/MemGUI-3K
language:
- en
library_name: transformers
license: apache-2.0
pipeline_tag: image-text-to-text
tags:
- qwen3_vl
- gui-agent
- mobile-gui
- android
- memory
- context-management
- conact
- memgui-agent
- long-horizon
model-index:
- name: MemGUI-8B-SFT
  results:
  - task:
      type: image-text-to-text
      name: Long-horizon mobile GUI control
    dataset:
      name: MemGUI-Bench
      type: lgy0404/MemGUI-3K
    metrics:
    - type: pass_at_1
      value: 23.4
      name: Pass@1
    - type: pass_at_3
      value: 35.9
      name: Pass@3
    - type: irr
      value: 30.2
      name: Information Retention Rate
  - task:
      type: image-text-to-text
      name: Out-of-distribution mobile GUI control
    dataset:
      name: MobileWorld GUI-Only
      type: mobileworld-gui-only
    metrics:
    - type: success_rate
      value: 17.9
      name: Success Rate
---

# MemGUI-8B-SFT

[**Project Page**](https://memgui-agent.github.io/) | [**Paper**](https://huggingface.co/papers/2606.19926) | [**Code**](https://github.com/kwai/MemGUI-Agent)

MemGUI-8B-SFT is an 8B MemGUI-Agent model trained from Qwen3-VL-8B-Instruct on
MemGUI-3K. It is designed for long-horizon mobile GUI control with proactive
context management.

The model follows the ConAct Context-as-Action protocol. At each step, it
produces a structured response with reasoning, history folding, a UI or memory
tool call, a grounded UI observation, and the next action intent. This allows
the agent to manage three context fields while acting: Folded Action History,
Folded UI State, and Recent Step Record.

## Model Details

- **Model type:** multimodal mobile GUI agent
- **Base model:** `Qwen/Qwen3-VL-8B-Instruct`
- **Training data:** `lgy0404/MemGUI-3K`
- **Training recipe:** supervised fine-tuning with ms-swift
- **Output protocol:** ConAct 5-part structured output
- **License:** Apache 2.0

## Intended Use

MemGUI-8B-SFT is intended for research on mobile GUI agents, long-horizon GUI
control, context management, UI memory, and history folding. It can be used as
an action policy in mobile GUI environments that provide screenshots and execute
structured tool calls.

This model is not a general-purpose chatbot. It expects the MemGUI-Agent system
prompt, a screenshot, and a structured mobile GUI context state.

## Input and Output Format

The model expects a multimodal conversation with:

- a system prompt defining the MemGUI-Agent tools and response format,
- a user message containing `<image>` plus the task goal and structured context,
- one screenshot image.

The assistant response follows this order:

```xml
<thinking>...</thinking>
<folding>{"range": [start_step, current_step], "summary": "..."}</folding>
<tool_call>{"name": "mobile_use", "arguments": {...}}</tool_call>
<ui_observation>...</ui_observation>
<action_intent>...</action_intent>
```

For the first step of a trajectory, `<folding>` is omitted because there is no
previous step to fold.

## Evaluation

| Benchmark            |       Metric | Score |
| -------------------- | -----------: | ----: |
| MemGUI-Bench         |       Pass@1 |  23.4 |
| MemGUI-Bench         |       Pass@3 |  35.9 |
| MemGUI-Bench         |          IRR |  30.2 |
| MobileWorld GUI-Only | Success Rate |  17.9 |

On MemGUI-Bench, MemGUI-8B-SFT improves over the Qwen3-VL-8B-Instruct baseline
and achieves the best open-data 8B performance reported in our experiments.
On MobileWorld GUI-Only, it transfers beyond the source benchmark and reaches
17.9% success rate.

## Dataset

MemGUI-3K contains 2,956 successful mobile GUI trajectories and 64,430
reasonable step-level training samples with ConAct annotations. The dataset
includes full trajectories, screenshots, step-level reasonableness annotations,
and multimodal training files.

Dataset page: https://huggingface.co/datasets/lgy0404/MemGUI-3K

## Citation

```bibtex
@article{memguiagent2026,
  title = {MemGUI-Agent: An End-to-End Long-Horizon Mobile GUI Agent with Proactive Context Management},
  author = {Guangyi Liu and Gao Wu and Congxiao Liu and Pengxiang Zhao and Liang Liu and Mading Li and Qi Zhang and Mengyan Wang and Liang Guo and Yong Liu},
  year = {2026},
  journal = {arXiv preprint arXiv:2606.19926}
}
```