Impact of large language models on peer review opinions from a fine-grained perspective: Evidence from top conference proceedings in AI

Peer review is a critical quality control mechanism in the academic research and publication process. Its primary purpose is to ensure the rigor and credibility of academic research, assist authors in improving their work, and identify potential errors and shortcomings. However, in recent years, the peer review mechanism has faced widespread criticism due to the surge in paper submissions and the shortage of domain experts qualified to serve as reviewers, particularly at top artificial intelligence conferences. Current peer review processes face several challenges, including bias, variability in review quality, unclear reviewer motivations, and imperfect review mechanisms. As submission volumes continue to rise, these issues are becoming increasingly pronounced. Some researchers have sought to mitigate these problems by enhancing fairness, reducing biases among novice reviewers, calibrating noisy peer review ratings, and improving mechanisms for matching papers with reviewers’ expertise. Other studies have explored the use of natural language processing techniques to support or refine the peer review process. These studies introduce the possibility of leveraging artificial intelligence to assist overburdened scientists in the peer review process. While these technologies may aid reviewers to some extent, their impact on the peer review process still requires further study.

In recent years, the impressive capabilities demonstrated by large language models (LLMs) have sparked extensive research and discussion within the academic community. At the same time, concerns have emerged within the academic community about the potential erosion of peer review by LLMs. Researchers have also begun to study and analyze the application and impact of LLMs in the peer review process. For example, Liang et al. not only evaluated the effectiveness of GPT-4 in generating scientific feedback but also proposed a method to estimate the extent of LLMs usage in peer review texts. They found that some reviews from recent AI conferences may have been modified by LLMs. Latona et al. investigated the prevalence and impact of LLM-assisted peer reviews at the ICLR 2024, finding that LLM-assisted reviews significantly influence review scores and submission acceptance rates. While these preliminary studies indicate that LLMs have begun to affect peer review, whether they are altering the core functions of peer review remains underexplored. As LLMs become increasingly integrated into scholarly workflows, it is therefore important to analyze their impacts on peer review from multiple perspectives, including linguistic patterns, evaluation aspects, and the influence of reviewer recommendations. Understanding these effects can help refine peer review practices, ensure fairness and transparency, and provide actionable insights for adapting to the evolving academic landscape. Notably, major conferences such as NeurIPS have not yet established explicit policies on whether reviewers may use LLMs to assist in writing their reports. This policy ambiguity underscores the importance of examining how LLM assistance may already be shaping the linguistic and evaluative characteristics of peer review texts.

This study aims to explore the impact of large language models (LLMs) on the linguistic complexity and content-level expression of peer review texts. Specifically, the research focuses on the following three core questions:

1. How has the emergence of LLMs affected the linguistic complexity and aspect-level content expression of peer review texts?

2. In comparison to non-LLM-assisted reviews, which evaluation aspects are more prominently emphasized in LLM-assisted reviews?

3. How do the evaluation aspects emphasized in LLM-assisted reviews relate to reviewers’ scoring and confidence levels?

Addressing these questions is crucial for understanding the role of LLMs in the academic review process and how to improve review practices.

The core innovations of this study include:

• �� Fine-Grained Analysis: This is the first study to systematically analyze the impact of large language models on peer review texts from a fine-grained perspective, distinguishing it from previous studies that only focused on overall text changes.

• �� Methodological Innovation: The study introduces a novel analytical framework that combines maximum likelihood estimation and automatic annotation techniques to identify peer review reports potentially modified or generated by LLMs.

• �� Data Analysis: By analyzing peer review texts from ICLR and NeurIPS conferences, the study reveals the impact of LLMs on review text length, linguistic complexity, and evaluative focus.

These innovations provide new perspectives for understanding the role of large language models in academic communication.

The study employs a detailed analytical approach, with the following steps:

• �� Data Collection and Processing: Selected ICLR and NeurIPS as data sources, using the OpenReview platform to obtain peer review text data.

• �� Review Sentence Aspect Identification: Based on the study by Yuan et al., a pre-trained aspect identification model is used to automatically annotate review sentences, identifying eight evaluative aspects.

• �� LLM-Assisted Peer Review Text Detection: Utilizes the maximum likelihood estimation model designed by Liang et al., combined with a predefined terminology dictionary, to detect peer review texts potentially assisted by LLMs.

• �� Lexical and Syntactic Complexity Analysis: Uses TAALES and TAASSC tools to calculate the lexical and syntactic complexity of review texts, analyzing their trends over time.

The experimental design includes the following aspects:

• �� Datasets: Selected peer review text data from ICLR and NeurIPS conferences, covering multiple years to analyze changes before and after the emergence of large language models.

• �� Baselines: Compared with non-LLM-assisted peer review texts to analyze the impact of LLM assistance on review texts.

• �� Evaluation Metrics: Includes review text length, lexical complexity, syntactic complexity, and the distribution of evaluative aspects.

• �� Hyperparameters: Adjusted parameters in the maximum likelihood estimation model to improve the accuracy of LLM-assisted text detection.

The study results indicate:

• �� Following the emergence of LLMs, peer review texts at ICLR and NeurIPS have become longer and more fluent, particularly among reviewers with lower confidence scores, where text length and fluency significantly increased.

• �� In LLM-assisted review reports, there is an increased emphasis on summaries and surface-level clarity, while attention to originality, replicability, and nuanced critical reasoning has declined.

• �� LLM-assisted review reports have a modest positive influence on the informativeness of recommendations. These findings reveal the profound impact of large language models on the academic peer review process, particularly in terms of changes in linguistic expression and evaluative dimensions.

The application scenarios of this study include:

• �� Academic Review: Provides references for improving review practices at academic conferences and journals, especially in terms of linguistic and evaluative dimensions of review texts.

• �� Education Sector: Offers insights into the role of large language models in academic communication, helping students and researchers better understand and utilize these technologies.

• �� AI Research: Provides empirical data on the application and impact of large language models in natural language processing and generation for AI researchers.

Despite revealing the impact of large language models on peer review, the study has some limitations:

• �� Data Limitation: The study primarily relies on publicly available data from ICLR and NeurIPS, which may not be applicable to conferences in other fields.

• �� Detection Model Accuracy: The accuracy of the maximum likelihood estimation method depends on the quality of the training data, which may lead to misjudgments.

• �� Subjective Bias: The potential impact of large language models on reviewer biases was not explored in depth. Future research could extend to peer reviews in other academic fields to explore the impact of large language models on review processes across different domains. Additionally, more precise detection models could be developed to enhance the identification of content generated by large language models.