UKPLab/ProReviewer-8B
ProReviewer-8B by UKPLab is an 8 billion parameter, Qwen3-based causal language model specifically fine-tuned for scientific peer review. Utilizing Group Relative Policy Optimization (GRPO) and trained on ICLR 2025 papers, it excels at generating high-quality, evidence-based reviews. This model is designed as the backbone for the ProReviewer agent, enabling structured investigation of scientific papers rather than passive generation, and supports a 32768 token context length.
Loading preview...
ProReviewer-8B: An RL-Trained Scientific Peer Review Model
ProReviewer-8B, developed by UKPLab, is an 8 billion parameter model built upon the Qwen3-8B architecture. It is uniquely fine-tuned using Group Relative Policy Optimization (GRPO) with step-level advantages to produce high-quality, evidence-based peer reviews of scientific papers. The model's training data consists of ICLR 2025 papers from the UKPLab/ProReviewer-Dataset.
Key Capabilities
- Specialized Peer Review: Generates structured, evidence-based reviews for scientific papers.
- RL-Trained: Leverages GRPO for enhanced review quality and reasoning.
- Agent Backbone: Designed as the core model for the ProReviewer agent, which performs R1-style structured investigation.
- Extended Context: Supports a context length of 32768 tokens, suitable for processing full scientific papers.
Recommended Usage
This model is primarily intended for use within the ProReviewer agent framework for comprehensive scientific paper analysis. It can also be served via vLLM or integrated directly using the Hugging Face Transformers library for specific tasks requiring its specialized review capabilities. Further details are available in the associated research paper.