---
language:
- en
library_name: transformers
tags:
- gui-grounding
- vision-language-model
- self-distillation
- qwen3.5
- gui-rd
license: apache-2.0
base_model: Qwen/Qwen3.5-9B
pipeline_tag: image-text-to-text
---

# GUI-RD-9B

This is the GUI-RD checkpoint for **Trust the Right Teacher: Quality-Aware Self-Distillation for GUI Grounding**.

GUI-RD stands for **GUI Reliable Distillation**. The checkpoint corresponds to the paper main setting: failed-token gate `0.5`, scaling coefficient `3`, and teacher top1-probability scaling enabled.

- Paper: https://arxiv.org/abs/2606.18101
- Hugging Face paper page: https://huggingface.co/papers/2606.18101
- Code: https://github.com/HarveyYellow/GUI-RD

## Intended Use

This model is intended for GUI grounding research and evaluation. It takes a GUI screenshot and a natural-language instruction, then predicts the target screen coordinate.

## Loading

```python
from transformers import AutoModelForMultimodalLM, AutoProcessor
import torch

model_id = "JingyuanHuang/GUI-RD-9B"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForMultimodalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
```

Depending on your installed Transformers version, the concrete auto-model class for Qwen3.5 may differ. For older Transformers releases, use `torch_dtype=torch.bfloat16` instead of `dtype=torch.bfloat16`. The repository provides standard Transformers config, tokenizer, processor, and safetensors weights.

## Citation

```bibtex
@misc{huang2026trustrightteacherqualityaware,
      title={Trust the Right Teacher: Quality-Aware Self-Distillation for GUI Grounding},
      author={Jingyuan Huang and Zuming Huang and Yucheng Shi and Tianze Yang and Xiaoming Zhai and Wei Chu and Ninghao Liu},
      year={2026},
      eprint={2606.18101},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2606.18101},
}
```