CIKM 2025
Proceedings of the 34th ACM International Conference on Information and Knowledge Management, 1572–1581.
An aspect-guided, multi-level perturbation framework probes the reliability of large language models in automated peer review. Controlled changes to papers, reviews, and rebuttals test responses to contribution, soundness, presentation, tone, and completeness. The analysis identifies recurring weaknesses in evaluating methodological flaws, strong rejections, incomplete rebuttals, and poor critiques, informing the design of human–AI peer-review partnerships.
@inproceedings{10.1145/3746252.3761274,
author = {Li, Jiatao and Li, Yanheng and Hu, Xinyu and Gao, Mingqi and Wan, Xiaojun},
title = {Where Do LLMs Go Wrong? Diagnosing Automated Peer Review via Aspect-Guided Multi-Level Perturbation},
year = {2025},
isbn = {9798400720406},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3746252.3761274},
doi = {10.1145/3746252.3761274},
booktitle = {Proceedings of the 34th ACM International Conference on Information and Knowledge Management},
pages = {1572–1581},
numpages = {10},
location = {Seoul, Republic of Korea},
series = {CIKM '25}
}