---
library_name: transformers
base_model: Qwen/Qwen3-8B
model_type: qwen3
pipeline_tag: text-generation
license: mit
language:
  - en
tags:
  - peer-review
  - scientific-papers
  - GRPO
  - reinforcement-learning
  - paper-review
datasets:
  - UKPLab/ProReviewer-Dataset
citation: |
  @article{fang2026passive,
    title={From Passive Generation to Investigation: A Proactive Scientific Peer Review Agent},
    author={Fang, Haishuo and Feng, Yue and Gurevych, Iryna},
    journal={arXiv preprint arXiv:2606.13349},
    year={2026}
  }
---

# ProReviewer-8B

An RL-trained scientific peer review model based on [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B). ProReviewer-8B is fine-tuned using Group Relative Policy Optimization (GRPO) to produce high-quality, evidence-based peer reviews of scientific papers.

## Model Description

ProReviewer-8B is the backbone model for the **ProReviewer** agent, an R1-style reasoning agent that reviews scientific papers through structured investigation rather than passive generation. The model was trained with a multi-stage curriculum:


### Training Details

| Parameter | Value |
|-----------|-------|
| Base model | Qwen/Qwen3-8B |
| Training method | SFT+ GRPO with step-level advantages |
| Training data | ICLR 2025 papers ([UKPLab/ProReviewer-Dataset](https://huggingface.co/datasets/UKPLab/ProReviewer-Dataset)) |
| Architecture | Qwen3ForCausalLM |
| Parameters | 8B |
| Precision | bfloat16 |

## Usage

### With vLLM

```bash
vllm serve UKPLab/ProReviewer-8B --max-model-len 32768 --dtype bfloat16
```

### With the ProReviewer Agent

The recommended way to use this model is through the ProReviewer agent framework in the [ProReviewer](https://github.com/UKPLab/arxiv2026-ProReviewer):

```python
from reviewer.evaluation import run_inference

paper = {
    "paper_id": "example",
    "paper_content": "# Paper Title\n\nAbstract: ...",
    "human_avg_score": 5.0,
}

# Option 1: Use a config name from config.toml (model served via vLLM)
result = await run_inference(paper, model="proreviewer-8B")

# Option 2: Use a local path (loads model directly via vLLM)
result = await run_inference(paper, model="/path/to/ProReviewer-8B")
```


### With Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("UKPLab/ProReviewer-8B", torch_dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained("UKPLab/ProReviewer-8B")
```

## Associated Resources

- **Paper**: [From Passive Generation to Investigation: A Proactive Scientific Peer Review Agent](https://arxiv.org/abs/2606.13349)
- **Code**: [UKPLab/arxiv2026-ProReviewer](https://github.com/UKPLab/arxiv2026-ProReviewer)
- **Dataset**: [UKPLab/ProReviewer-Dataset](https://huggingface.co/datasets/UKPLab/ProReviewer-Dataset)

## Citation

```bibtex
@article{fang2026passive,
  title={From Passive Generation to Investigation: A Proactive Scientific Peer Review Agent},
  author={Fang, Haishuo and Feng, Yue and Gurevych, Iryna},
  journal={arXiv preprint arXiv:2606.13349},
  year={2026}
}
```

## License

This model is released under the [MIT License](https://opensource.org/licenses/MIT).