🤖 AI

PRAIB: Peer Review AI Benchmark of Behaviour of LLM-Assisted Reviewing

This paper introduces PRAIB, a novel benchmark framework and large-scale empirical study analyzing 11,000 machine-generated reviews against human feedback to reveal significant behavioral divergences—such as positive bias, overconfidence, and missed atomic weaknesses—thereby providing a diagnostic tool to determine the current reliability and limitations of LLMs in the peer review process.

Original authors: Krzysztof Żurawicki, Julia Farganus, Arkadiusz Gaweł, Mateusz Bystroński, Tomasz Jan Kajdanowicz

Published 2026-05-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Krzysztof Żurawicki, Julia Farganus, Arkadiusz Gaweł, Mateusz Bystroński, Tomasz Jan Kajdanowicz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive, high-stakes talent show where thousands of scientists submit their best work to be judged by a panel of experts. For decades, this "peer review" process has been done entirely by humans. But recently, a new kind of judge has entered the arena: Artificial Intelligence (AI).

The paper you are asking about, PRAIB, is like a giant quality control test designed to answer one big question: Is the AI actually "thinking" like a human judge, or is it just a fancy robot that sounds smart but doesn't really understand the work?

Here is the breakdown of what the researchers found, using some everyday analogies.

1. The Setup: The "Lazy" Robot vs. The Human Expert

The researchers took 1,000 real scientific papers from top conferences (ICLR and NeurIPS) and asked five different AI models to write reviews for them. They then compared these AI reviews against the actual reviews written by human experts.

Think of it like asking a group of students to write a book report.

The Human: Reads the book, highlights specific paragraphs, checks the math, and writes a thoughtful critique.
The AI: Reads the book (or tries to), and then writes a review that looks like a book report.

2. The Main Finding: The AI is "Over-Confident" and "Over-Wordy"

The study found that while AI reviews look impressive on the surface, they behave very differently from humans.

The "Verbose" Problem: The AI reviews were often much longer and more complicated than human reviews.
- Analogy: Imagine a human chef saying, "The soup needs more salt." The AI chef, however, writes a three-page essay explaining the history of salt, the chemistry of sodium, and the philosophy of seasoning, but still forgets to actually say the soup needs salt. The AI is trying to sound smart by using big words and long sentences, making it harder to read (lower "readability scores").
The "Confidence" Problem: The AI was extremely confident in its opinions, even when it was wrong.
- Analogy: If a human judge is unsure about a paper, they might say, "I'm not 100% sure, but I think..." The AI, however, acts like it knows everything, giving a perfect score with 100% certainty, even if it missed the main point. It's like a student who guesses on a test but marks the answer sheet with a giant, confident checkmark.
The "Positive Bias": The AI tended to give higher scores than humans. It was nicer than the human judges, often overlooking the paper's flaws.

3. The "Specificity" Test: Does it actually look at the paper?

This was the most critical part of the test. The researchers checked if the AI actually looked at the specific details of the paper, like Figure 4, Equation 12, or Table 3.

The Human: "In Figure 4, the graph shows a spike that doesn't make sense."
The AI: "The figures are generally well presented." (Vague and generic).

The study found that many AI models failed to point to specific parts of the paper. They were like a tourist who looks at a museum map and says, "There are lots of paintings here," without ever stopping to look at a specific painting.

The Exception: One specific AI model (called OpenReviewer) was trained specifically on review data. It behaved much more like a human, pointing to specific figures and writing in a style that felt natural. This proves that if you train an AI specifically for the job, it can do a better job than a general-purpose AI.

4. The "Math" Problem

Scientific papers are full of math. The researchers checked if the AI actually engaged with the equations or just talked around them.

Some AI models were surprisingly good at finding math and discussing it.
However, others completely ignored the math, treating a complex physics paper like a casual blog post.

5. The "Fake Citation" Issue

A major red flag found in the study was that the AI models often made up references.

Analogy: If a human writer says, "As Smith (2020) proved...", they are quoting a real book. The AI, however, would sometimes say, "As Smith (2020) proved..." but "Smith (2020)" didn't actually exist. It was a hallucination—a confident lie. The study found that most AI models failed to provide real, verifiable citations.

The Conclusion: AI is a "Co-Pilot," Not the "Pilot"

The paper concludes that we cannot simply replace human reviewers with AI right now.

The AI is good at: Writing long, polite, and structured text. It can help draft reviews or check for basic grammar.
The AI is bad at: Being honest about its confidence, finding specific flaws in the data, and providing real, verifiable evidence.

The Final Verdict:
Think of the AI as a very well-spoken intern. It can write a beautiful report, but it might miss the crucial detail that the engine is broken because it's too busy using big words. The human expert is still needed to be the senior manager who checks the facts, verifies the citations, and makes the final decision.

The researchers created a "scorecard" (PRAIB) to help future developers build better AI tools that don't just sound smart, but actually act like smart, careful human reviewers.

Technical Summary: PRAIB: Peer Review AI Benchmark of Behaviour of LLM-Assisted Reviewing

Problem Statement

The increasing volume of scientific submissions has driven the exploration of Large Language Models (LLMs) to support and augment the peer review process. However, it remains unclear whether LLMs engage with scientific manuscripts in a manner consistent with human reviewers or if they merely generate text that superficially resembles a review. Existing evaluation approaches often focus on numerical rating prediction, "LLM-as-a-Judge" pipelines, or surface-level text statistics (e.g., helpfulness, verifiability). These methods frequently overlook the critical dimension of human alignment: whether the behavioral patterns, specificity, and engagement style of an AI-generated review mirror those of human experts. There is a need for a diagnostic framework to identify systematic behavioral divergences between human and machine reviewers, particularly as the use of generative AI in reviewing accelerates.

Methodology

The authors introduce the Peer Review AI Benchmark (PRAIB), a multi-dimensional framework designed to measure the specificity, style, and engagement behavior of LLM-assisted reviews. The methodology involves a large-scale empirical study comparing machine-generated reviews against human feedback.

Data Acquisition and Preparation

Dataset: The study utilizes a dataset of 11,000 reviews generated for 1,000 papers from the ICLR and NeurIPS conferences spanning the years 2021–2025.
Models Evaluated: Five models were tested: three open-source models (DeepSeek-V3, Gemma-3-12B-it, Qwen3.5-9B), one domain-specialized model (OpenReviewer-8B), and one proprietary frontier model (GPT-5-2025-08-07).
Prompting Strategies: Models were evaluated using two distinct prompting templates: a "lazy" baseline (conference-agnostic) and detailed conference-specific templates (aligned with official ICLR/NeurIPS guidelines).
Preprocessing: Manuscripts were converted from PDF to Markdown using DeepSeek-OCR to preserve mathematical fidelity. Reviews were truncated at 50,000 tokens to focus on the main body.

Evaluation Framework (PRAIB)

The framework quantifies behavioral patterns across several axes:

Textual Complexity and Readability:
- Metrics include Token Count (TKC), Type-Token Ratio (TTR), Flesch Reading Ease (FRE), and Flesch-Kincaid Grade Level (F-K Grade).
- These measure linguistic richness, sentence complexity, and readability compared to human baselines.
Content Engagement (Specificity):
- Cross-References (Xref.): Extraction of references to specific manuscript elements (figures, tables, equations, theorems) using regular expressions.
- Mathematical Content Detection: A secondary LLM (Qwen3.5-8B) classifies whether a review substantively engages with mathematical content (e.g., citing specific equations, analyzing proofs). This was validated against human annotations with 96.6% accuracy.
- External Citations: Verification of citations against external bibliographic databases (Crossref, OpenAlex, DBLP) to detect hallucinations and ensure factuality.
Reviewer Agreement:
- Krippendorff's $\alpha$ : Used to measure inter-annotator agreement on ordinal rating scales, accounting for chance agreement and the ordered nature of scores.
- Conditional Math Detection: Measures $P(+|+)$ (LLM engages with math given human did) and $P(+|-)$ (LLM engages with math given human did not) to assess alignment in technical focus.
Information Coverage:
- Atomic Extraction: Reviews are decomposed into "atomic" strengths and weaknesses (self-contained units of information).
- Coverage Metrics: Atomic Recall (fraction of human points covered by LLM), Precision, Mapping Density, and Compaction Ratio are calculated to determine if LLMs cover the same technical ground as humans.

Key Results

1. Divergence in Stylistic and Behavioral Patterns

Readability: LLM-generated reviews are significantly more complex and difficult to read than human reviews. Flesch Reading Ease scores for LLMs (13.7–21.3) are far below the human baseline (39.1), and F-K Grade levels are higher (14.5–15.7 vs. 12.1).
Length: LLMs tend to produce longer reviews. GPT-5 averages 1,449 tokens (2.7 $\times$ human mean), while even more restrained models like DeepSeek exceed human length by 20%.
Vocabulary: Type-Token Ratios (TTR) are lower for LLMs, indicating reduced vocabulary diversity and repetitive phrasing.

2. Specificity and Engagement

Cross-References: While some models (Qwen, GPT-5) generate high numbers of cross-references, others (DeepSeek, Gemma) produce very few. Notably, OpenReviewer (fine-tuned) closely matches human cross-reference counts (2.77 vs. 2.26).
Mathematical Engagement: Frontier models (GPT-5, Qwen) show high rates of mathematical engagement (75.8% and 68.6%), exceeding the human rate (42.7%). However, smaller models (DeepSeek, Gemma) fall below human rates.
Citation Failure: Almost all general-purpose models failed to generate valid, verifiable citations. Only OpenReviewer produced some verifiable citations (33% verified), similar to the human rate (35%).

3. Rating Alignment and Bias

Agreement: GPT-5 achieved the highest agreement with human panels ( $\alpha_r \approx 0.36$ ), matching or slightly exceeding the human-only baseline. DeepSeek showed the weakest alignment ( $\alpha_r \approx 0.14$ ).
Bias: LLM ratings exhibit a positive bias (tendency to give higher scores) and low variance. They are less variable than human ratings and often cluster at the upper end of the scale.
Confidence: LLMs display overconfidence, rarely assigning low confidence scores (1 or 2) which humans use frequently.

4. Information Coverage

Volume vs. Quality: LLMs consistently generate more atomic strengths and weaknesses than humans. GPT-5 produces nearly triple the human baseline for weaknesses.
Coverage: While LLMs cover a high fraction of human-identified strengths (Atomic Recall $\approx$ 0.89–0.94), coverage of weaknesses is lower (0.70–0.93).
Pattern: Frontier models (GPT-5, Qwen) tend to decompose human points into finer sub-points (high mapping density), whereas OpenReviewer maintains a more one-to-one mapping.

Significance and Claims

The paper claims that PRAIB provides a diagnostic tool for the community to identify which aspects of the review process LLMs can reliably support today and which require further development.

Behavioral Divergence: The study demonstrates that LLMs do not merely mimic human reviews; they exhibit distinct behavioral patterns, including excessive verbosity, reduced vocabulary diversity, and a tendency to overlook atomic weaknesses noted by humans.
Limitations of Current Models: General-purpose models are currently unsuitable as wholesale replacements for human expertise due to failures in citation factuality, positive scoring bias, and inconsistent engagement with specific manuscript elements.
Potential for Augmentation: Specific models, particularly those fine-tuned for the domain (OpenReviewer) or frontier models with strong mathematical engagement (GPT-5), show promise in augmenting human-led technical reviews, particularly in mathematical analysis and identifying formal structures.
Future Directions: The authors propose a synergistic pipeline where frontier models handle mathematical analysis, fine-tuned models assist with stylistic drafting, and human reviewers retain final qualitative judgment and citation verification.

The paper concludes that while LLMs are accelerating the review process, their behavioral deviations from human norms necessitate careful evaluation and hybrid human-AI workflows rather than full automation.