This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
This paper proposes ReviewRL, a reinforcement learning-based automated paper review system, to improve the peer review process, which is struggling due to the increasing volume of paper submissions and reviewer fatigue. ReviewRL combines an ArXiv-MCP search-based context generation pipeline that integrates relevant scientific literature, supervised learning fine-tuning to establish baseline reviewing skills, and a reinforcement learning procedure that uses a compound reward function to improve review quality and rating accuracy. Experimental results on ICLR 2025 papers demonstrate that ReviewRL outperforms existing methods in both rule-based metrics and model-based quality assessment. This publication will be made publicly available on GitHub.
Takeaways, Limitations
•
Takeaways:
◦
We present the potential of an automated paper review system using reinforcement learning.
◦
It improves the limitations of existing automated review systems, such as factual accuracy, evaluation consistency, and analysis depth.
◦
Improve the quality of your review by leveraging relevant scientific literature.
◦
It outperforms existing methods in rule-based metrics and model-based quality assessment.
•
Limitations:
◦
Only experimental results for the ICLR 2025 paper are presented, so further verification of generalizability is needed.
◦
There is a lack of detailed explanation of the design and optimization of compound reward functions.
◦
After the release on GitHub, we need to evaluate its actual usability and effectiveness.
◦
Further research is needed to determine whether it can fully replace the expertise and insight of human judges.
