ReviewGuard: Automated Peer Review Detection

• ReviewGuard is an automated, LLM-based system that identifies and categorizes deficient peer reviews to enhance academic evaluation.
• It employs a four-stage pipeline—data acquisition, GPT-4.1 annotation with human validation, synthetic augmentation, and fine-tuned modeling—to ensure robust detection.
• The framework demonstrates improved detection metrics and supports policy development for responsible AI integration in scholarly review processes.

ReviewGuard is an automated, LLM–driven system for detecting and categorizing deficient peer reviews, with an emphasis on strengthening academic integrity in the presence of escalating review volumes and widespread AI involvement. The framework combines high-quality peer review datasets, rigorous annotation protocols (mixing GPT-4.1 and human verification), targeted data augmentation to resolve label imbalance, and fine-tuned classification models. Comprehensive analyses demonstrate systematic distinctions between sufficient and deficient reviews, an acute rise in AI-generated content, and substantial improvements in deficient review detection through synthetic augmentation. ReviewGuard is the first system of its kind to automate and enhance deficient peer review detection at scale, providing robust tools and benchmarks for human-AI collaboration and policy development in scholarly evaluation (Zhang et al., 18 Oct 2025).

1. Motivation and Problem Definition

Peer review underpins scientific quality control, but increasing submission rates and the adoption of LLMs for review writing have led to new challenges—namely, the production and proliferation of deficient reviews. Deficient reviews, whether written by humans or AI systems, are characterized by lack of commitment, poor constructiveness, or insufficient subject matter understanding. Such reviews not only jeopardize individual publication decisions but also threaten the peer review system’s reliability and fairness. The unchecked spread of LLM-generated reviews further complicates detection, masking deficiencies behind plausible but generic text. ReviewGuard is introduced to address these emerging, systematic threats to the integrity and trustworthiness of scholarly assessment.

2. Data Collection and Annotation Pipeline

ReviewGuard implements a four-stage LLM-centric pipeline:

(a) Data Acquisition: The system leverages OpenReview data from ICLR and NeurIPS, with a focus on papers manifesting high reviewer disagreement. For each paper, a consensus score is assigned (average over reviewer ratings, omitting extremes), and cases exceeding a preset diff value (e.g., ≥ 3) between a review and consensus are preferentially sampled to enrich the corpus for likely deficiencies.

(b) Annotation & Labeling: Reviews are labeled using GPT-4.1, following strict definitions: reviews are split into “sufficient” (meeting all standards for commitment, constructiveness, and domain-expertise) and “deficient.” Deficient reviews are further classified into subtypes—superficial, non-constructive, cursory, excessively harsh, uninformed, or biased. This automated process is followed by expert human validation (agreement measured by Cohen’s and Fleiss’s kappa).

(c) Synthetic Data Augmentation: To counter class imbalance (deficient reviews being a minority), ReviewGuard uses prompt-driven LLMs to generate new samples. Definitions for each subtype serve as prompt templates, producing extra deficient and sufficient reviews for each paper. This synthetic corpus (totaling 46,438 synthetic samples) is merged with originals (24,657 real reviews) for robust model training.

(d) Model Fine-Tuning: Both encoder-only transformers (e.g., BERT-base-uncased, SciBERT, RoBERTa) and open-source LLMs (such as LLaMa 3.1-8B-Instruct, Qwen 3-8B) are fine-tuned using the augmented dataset. Open-source LLMs are adapted via parameter-efficient fine-tuning (LoRA) with r = 8, α = 32, dropout = 0.05.

3. Linguistic and Structural Analysis of Review Quality

A comprehensive feature analysis reveals marked differences between sufficient reviews (SR) and deficient reviews (DR):

• Ratings: SRs have higher mean reviewer ratings (5.37) than DRs (3.74). There is a statistically significant positive correlation (Spearman’s ρ ≈ 0.256, p < 0.001) between review sufficiency and rating score.
• Structural Complexity: SRs employ more sentences (24.61 vs. 19.43), tokens (425.47 vs. 318.44), and demonstrate higher readability scores (Linsear Write, 14.02 vs. 12.68).
• Sentiment: 90% of SRs are neutral, whereas DRs show a 4–5× increase in negative sentiment, signifying more emotionally charged or biased feedback.
• Confidence: DRs surprisingly report higher reviewer self-confidence, despite poorer substance—indicating that confident assertions do not equate to high-quality assessment.

These insights confirm that DRs are typically shorter, more negative, structurally simpler, and paradoxically more confident, highlighting their potential risk for undermining scientific review processes.

4. Detection of AI-Generated Reviews

The analytical pipeline incorporates the Binoculars framework, leveraging models like Falcon-7B and Falcon-7B-Instruct to assess perplexity gaps and other statistical signals of LLM authorship. Temporal analyses indicate a marked surge in AI-written reviews post-ChatGPT release, impacting both sufficient and deficient review pools. This trend underscores the urgency of employing dedicated detection mechanisms and instituting transparent policies regarding AI involvement in peer review. This suggests heightened risk of hidden deficiencies as generic LLM-written text proliferates.

5. Model Training, Evaluation, and Results

ReviewGuard’s models are validated on both binary (SR vs. DR) and multi-label (subtype) classification tasks:

• Baseline (Real-Only): Encoder models (e.g., SciBERT) provide strong F1 baselines for binary sufficiency/deficiency.
• Effectiveness of Synthetic Augmentation: Introducing synthetic data alongside real reviews causes recall for Qwen 3-8B, for example, to rise from 0.5499 (R only) to 0.6653 (R+S); F1 improves from 0.5606 to 0.7073. Multi-label detection—which is inherently harder—also benefits across all tested models. This mixed-data approach enhances discovery of nuanced, low-frequency deficient patterns that are under-represented in native submissions.
• Model Adaptability: Both encoder-based and LLM-based approaches are effective when trained on the enriched (real + synthetic) corpus, striking a balance between classification performance and computational feasibility.

6. Implications for AI Governance and Peer Review

ReviewGuard has broad implications for peer review integrity, AI policy, and human-AI collaboration:

• Governance: Automated detection of deficient or AI-generated reviews enables conference organizers and journal editors to identify quality issues proactively, bolstering the reliability of the review process.
• Policy: The proliferation of LLM-written reviews highlights the necessity of explicit norms for disclosure, usage, and oversight of AI in peer review, as detection alone cannot prevent all risks of low-quality or misleading feedback.
• Human–AI Synergy: The framework shows that LLMs can be leveraged not only for augmentation and review generation, but also for annotation and QA, significantly reducing manual labeling effort and increasing annotation reliability. A plausible implication is that expanding human-AI collaboration for review assessment could create scalable, high-integrity workflows across disciplines.

Future enhancements may integrate reinforcement learning for still higher-quality synthetic review generation, and expand detection coverage to multimodal or multidisciplinary contexts.

7. Limitations and Prospective Extensions

ReviewGuard’s current applications are focused on ML/AI conference reviews, with expansion to broader academic domains, disciplines, and modalities as a future target. Deployment in real-time reviewing environments and integration with cross-journal governance could further amplify its impact. The framework’s reliance on robust annotation (automation combined with expert validation) and continual adaptation to evolving LLM usage patterns are likely to remain critical as peer review practices and technologies co-evolve.

---

ReviewGuard constitutes a comprehensive technical approach, encompassing dataset engineering, annotation, augmentation, and advanced modeling, to meet the emergent need for automated, high-fidelity detection of deficient peer reviews in increasingly LLM-mediated scholarly evaluation (Zhang et al., 18 Oct 2025).