Reproduce “EGTR-Review: Efficient Evidence-Grounded Scientific Peer Review Generation via Multi-Agent Teacher Distillation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinpeng Qiu, Wang Yihu, Zhifeng Liu, Xiaochen Wang, Jimin Wang
EGTR-Review, from Peking University, employs multi-agent teacher distillation to create scientific peer reviews that are both evidence-grounded and traceable. This method substantially reduces inference costs and time (e.g., 44 seconds per paper) while outperforming prior systems in review quality, particularly in evidence grounding and traceability.
Key insights
- Decomposing complex review generation into distinct stages of evidence gathering, verification, and synthesis, as performed by a multi-agent teacher, yields high-quality, traceable feedback.
- Distilling not just the final review output, but also the step-by-step reasoning process from a complex teacher, enables a smaller student model to retain sophisticated capabilities efficiently.
- An evidence-weighted objective function in distillation helps the student prioritize learning from robust, strongly supported judgments, improving the reliability and factual grounding of generated reviews.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.