Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ReviewGuard: Enhancing Deficient Peer Review Detection via LLM-Driven Data Augmentation

Haoxuan Zhang, Ruochi Li, Sarthak Shrestha, Shree Harshini Mamidala, Revanth Putta, Arka Krishan Aggarwal, Ting Xiao, Junhua Ding, Haihua Chen

TL;DR

ReviewGuard addresses the risk of deficient peer reviews in scalable, AI-augmented scholarly evaluation by building a four-stage LLM-driven framework that collects data from OpenReview (ICLR/NeurIPS), annotates with GPT-4.1 and human validation, augments with synthetic reviews, and trains encoder and open LLM detectors. The study produces a labeled corpus of 24,657 real reviews from 6,634 papers and 46,438 synthetic reviews, reveals that deficient reviews show lower ratings, less structural complexity, and more negative sentiment, and documents a post-ChatGPT rise in AI-generated reviews. Mixed training with real and synthetic data significantly boosts detection performance (e.g., Qwen 3-8B: recall 0.6653, F1 0.7073). These findings advance AI governance in peer review and provide practical, open resources for researchers and policymakers.

Abstract

Peer review serves as the gatekeeper of science, yet the surge in submissions and widespread adoption of large language models (LLMs) in scholarly evaluation present unprecedented challenges. While recent work has focused on using LLMs to improve review efficiency, unchecked deficient reviews from both human experts and AI systems threaten to systematically undermine academic integrity. To address this issue, we introduce ReviewGuard, an automated system for detecting and categorizing deficient reviews through a four-stage LLM-driven framework: data collection from ICLR and NeurIPS on OpenReview, GPT-4.1 annotation with human validation, synthetic data augmentation yielding 6,634 papers with 24,657 real and 46,438 synthetic reviews, and fine-tuning of encoder-based models and open-source LLMs. Feature analysis reveals that deficient reviews exhibit lower rating scores, higher self-reported confidence, reduced structural complexity, and more negative sentiment than sufficient reviews. AI-generated text detection shows dramatic increases in AI-authored reviews since ChatGPT's emergence. Mixed training with synthetic and real data substantially improves detection performance - for example, Qwen 3-8B achieves recall of 0.6653 and F1 of 0.7073, up from 0.5499 and 0.5606 respectively. This study presents the first LLM-driven system for detecting deficient peer reviews, providing evidence to inform AI governance in peer review. Code, prompts, and data are available at https://github.com/haoxuan-unt2024/ReviewGuard

ReviewGuard: Enhancing Deficient Peer Review Detection via LLM-Driven Data Augmentation

TL;DR

ReviewGuard addresses the risk of deficient peer reviews in scalable, AI-augmented scholarly evaluation by building a four-stage LLM-driven framework that collects data from OpenReview (ICLR/NeurIPS), annotates with GPT-4.1 and human validation, augments with synthetic reviews, and trains encoder and open LLM detectors. The study produces a labeled corpus of 24,657 real reviews from 6,634 papers and 46,438 synthetic reviews, reveals that deficient reviews show lower ratings, less structural complexity, and more negative sentiment, and documents a post-ChatGPT rise in AI-generated reviews. Mixed training with real and synthetic data significantly boosts detection performance (e.g., Qwen 3-8B: recall 0.6653, F1 0.7073). These findings advance AI governance in peer review and provide practical, open resources for researchers and policymakers.

Abstract

Peer review serves as the gatekeeper of science, yet the surge in submissions and widespread adoption of large language models (LLMs) in scholarly evaluation present unprecedented challenges. While recent work has focused on using LLMs to improve review efficiency, unchecked deficient reviews from both human experts and AI systems threaten to systematically undermine academic integrity. To address this issue, we introduce ReviewGuard, an automated system for detecting and categorizing deficient reviews through a four-stage LLM-driven framework: data collection from ICLR and NeurIPS on OpenReview, GPT-4.1 annotation with human validation, synthetic data augmentation yielding 6,634 papers with 24,657 real and 46,438 synthetic reviews, and fine-tuning of encoder-based models and open-source LLMs. Feature analysis reveals that deficient reviews exhibit lower rating scores, higher self-reported confidence, reduced structural complexity, and more negative sentiment than sufficient reviews. AI-generated text detection shows dramatic increases in AI-authored reviews since ChatGPT's emergence. Mixed training with synthetic and real data substantially improves detection performance - for example, Qwen 3-8B achieves recall of 0.6653 and F1 of 0.7073, up from 0.5499 and 0.5606 respectively. This study presents the first LLM-driven system for detecting deficient peer reviews, providing evidence to inform AI governance in peer review. Code, prompts, and data are available at https://github.com/haoxuan-unt2024/ReviewGuard
Paper Structure (27 sections, 4 figures, 6 tables)

This paper contains 27 sections, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Peer Review Acquisition
  5. Review Categorization and Annotation
  6. Definitions
  7. Annotation Protocol
  8. Synthetic Review Augmentation
  9. Feature Analysis for Deficient Reviews
  10. Textual Structure Analysis
  11. Textual Quality Analysis
  12. Development of Deficient Review Detection Model
  13. Experiments and Results
  14. Experimental Setup
  15. Dataset Preparation
  16. ...and 12 more sections

Figures (4)

  • Figure 1: The ReviewGuard pipeline for automated deficient review detection. The framework integrates four sequential stages: (1) targeted data acquisition from OpenReview focusing on conflicting reviews, (2) LLM-driven annotation with human validation and synthetic data augmentation to address class imbalance, (3) multi-dimensional feature analysis examining textual structure and quality, and (4) fine-tuning of encoder-based and open-source LLM detection models.
  • Figure 2: Distribution of review categories in ICLR and NeurIPS conferences.
  • Figure 3: Temporal analysis of AI-generated review detection using Binoculars framework in ICLR and NeurIPS conferences.
  • Figure 4: Distribution of semantic similarity scores between paper abstracts and reviews, for both real and synthetic reviews from the ICLR and NeurIPS datasets. Dashed lines indicate peak similarity for each group.