README.md

---
license: apache-2.0
base_model: memgui-agent-anonymous/MemGUI-8B-SFT
pipeline_tag: image-text-to-text
tags: [mobile-gui-agent, agent-memory, reinforcement-learning, qwen3-vl, conact]
---

# MemGUI-8B-RL

Anonymous release for the ICLR 2027 submission *FARPO: Folding-Aware Reward-Decoupled Policy Optimization for Long-Horizon Mobile GUI Agents* (project MemGUI-RL). Project page: https://memgui-rl-anonymous.github.io/

MemGUI-8B-RL is MemGUI-8B-SFT (Qwen3-VL-8B-Instruct supervised on MemGUI-3K) post-trained for 100
optimizer steps with **FARPO** (Folding-Aware Reward-Decoupled Policy Optimization, span-to-step ratio
rho = 9). The policy speaks the ConAct (Context-as-Action) interface of MemGUI-Agent: every response
contains a folding directive for its own history, an optional memory operation and the next GUI action.

| benchmark | MemGUI-8B-SFT | **MemGUI-8B-RL** |
| --- | --- | --- |
| MemGUI-Bench Pass@1 / Pass@3 / IRR (128 tasks) | 23.4 / 35.9 / 30.2 | **30.5 / 48.4 / 39.0** |
| MobileWorld GUI-only success rate (117 tasks, OOD) | 17.9 | **19.7** |

## Usage

The checkpoint is a standard `Qwen3VLForConditionalGeneration` model (weights in bf16, ~17.5 GB).
Use it with the ConAct system prompt and the MemGUI-Agent runtime unchanged:

```python
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration
model = Qwen3VLForConditionalGeneration.from_pretrained("memgui-rl/MemGUI-8B-RL", torch_dtype="bfloat16", device_map="auto")
processor = AutoProcessor.from_pretrained("memgui-rl/MemGUI-8B-RL")
```

Training code and launch scripts: https://github.com/memgui-rl-anonymous/MemGUI-RL
Training data: https://huggingface.co/datasets/memgui-rl/MemGUI-3K-Verl
Evaluation logs: https://huggingface.co/datasets/memgui-rl/MemGUI-RL-Eval