ReviewFact: Factual Errors in Peer-Review Comments

ReviewFact is the benchmark of the paper PremiseCheck: Efficient Fact-Checking of Human and AI Peer Reviews.

Dingjie Song¹, Zhiling Yan¹, Ziqian Zhang¹, Yuxuan Zhang^2,3, Lichao Sun^1
1Lehigh University, ²University of British Columbia, ³Vector Institute

(a) The task on a real ReviewFact comment that claims a baseline is missing although the paper reports it. (b) ReviewFact labels come from the expert annotations of three datasets. For PeerReview Bench, two authors independently re-check which errors are factual on 178 comments.

TL;DR

• Task. Given a paper and one comment from a review of it, decide whether the comment contains a factual error about the paper, that is, says something about the paper that the paper does not support.
• Labels. Every label starts from an expert judgment in an existing dataset: ReviewCritique (NLP), ReviewScore (machine learning) and PeerReview Bench (natural sciences). No new comments or papers were added.
• Reviewers. Comments come from human reviews in all three parts and from AI (LLM) reviews in the NLP and natural-science parts.
• Size. The full benchmark has 9,267 comments on 217 papers. This repository distributes the NLP and natural-science parts (9,051 comments on 178 papers). The machine-learning part (216 comments) is rebuilt locally from its source with a script.
• Metric. AUROC with factual errors as the class to detect, reported together with the cost and latency of the system.
• Use. Evaluation only. Do not train models on ReviewFact.

Dataset structure

Configs

config	file	rows	papers	errors	content
nlp (default)	data/nlp/test.jsonl	6,466	100	308	review comments, NLP (ReviewCritique)
nlp_papers	data/nlp/papers.jsonl	100			paper texts for nlp
science	data/science/test.jsonl	2,585	78	65	review comments, natural sciences (PeerReview Bench)
science_papers	data/science/papers.jsonl	78			paper texts for science
machine learning	not distributed	216	39	46	rebuilt from ReviewScore, see below

Every config has a single test split. There is no training split.

Comments by reviewer type (correct / error):

part	human-written	AI-written
NLP	5,783 / 278	375 / 30
Natural sciences	1,686 / 40	834 / 25
Machine learning	170 / 46	none

Comment fields (nlp, science)

field	type	description
id	string	unique comment id (prefix rc| for the NLP part, prb| for the natural-science part)
source	string	source dataset: rc (ReviewCritique) or prb (PeerReview Bench)
split	string	always test
paper_id	string	key into the matching papers config
comment	string	the review comment (a segment of a review in the NLP part, a review item in the natural-science part)
reviewer_type	string	human or ai (the review was written by a person or by an LLM)
label	string	error or correct
error_type	string or null	source error category for error rows (Misunderstanding, Neglect, Unstated statement, Inaccurate Summary); null for correct rows
label_origin	string	expert (label taken from the source's expert judgment) or expert+llm-filter (the expert judged the comment incorrect and an LLM decided from the expert's rationale that the error is factual, see Human verification)

Paper fields (nlp_papers, science_papers)

field	type	description
paper_id	string	paper id
title	string	paper title
text	string	full text of the paper. The systems in the paper read the first 24k tokens (o200k tokenizer).

Example

The comment shown in the figure, from a human review of LoRA: Low-Rank Adaptation of Large Language Models (paper_id rc71). The paper reports full fine-tuning as a baseline in Section 5.1, so the experts labeled the comment a Neglect error.

{
  "id": "rc|human|rc71|review#4|27",
  "source": "rc",
  "split": "test",
  "paper_id": "rc71",
  "comment": "The authors did not report any baselines outside of the efficient fine-tuning setting.",
  "reviewer_type": "human",
  "label": "error",
  "error_type": "Neglect",
  "label_origin": "expert"
}

Quick start

from datasets import load_dataset

nlp = load_dataset("songdj/ReviewFact", "nlp", split="test")                 # comments
nlp_papers = load_dataset("songdj/ReviewFact", "nlp_papers", split="test")   # paper texts
science = load_dataset("songdj/ReviewFact", "science", split="test")
science_papers = load_dataset("songdj/ReviewFact", "science_papers", split="test")

# join comments to papers on paper_id
papers = {p["paper_id"]: p for p in nlp_papers}
example = nlp[0]
paper_text = papers[example["paper_id"]]["text"]

# binary target: 1 = factual error
y = [int(label == "error") for label in nlp["label"]]

To score a system, compute AUROC over all comments of a part with error as the positive class, for example with sklearn.metrics.roc_auc_score(y, error_scores). The PremiseCheck code provides premisecheck eval, which also reports cost, latency and the paired paper-bootstrap difference to a reference run. To download the raw files instead:

hf download songdj/ReviewFact --repo-type dataset --local-dir data/ReviewFact

Task and labels

A comment contains a factual error when it says something about the paper that the paper does not support. ReviewFact adopts the four factual error types of ReviewCritique:

• Misunderstanding: the comment misreads the paper.
• Neglect: the comment claims that the paper lacks something it contains.
• Unstated statement: the comment asserts something with no basis in the paper.
• Inaccurate summary: the comment misstates the content of the paper.

Errors of judgment, such as whether a concern is important, are outside the task.

Each source dataset was built for a broader goal, so its annotations are mapped to one label and comments outside the task are left out:

part	source	annotators	correct	error	left out
NLP	ReviewCritique (Du et al., 2024)	40 NLP researchers, two per paper, disagreements resolved by an expert with area-chair experience	segment judged reliable	segment judged deficient with type Misunderstanding, Neglect, Unstated statement or Inaccurate Summary	other deficiencies, summary and strength segments, segments of fewer than four words
Natural sciences	PeerReview Bench (Kim et al., 2026)	45 scientists in the physical, biological and health sciences	comment judged correct	comment judged incorrect, and the error is factual (decided from the expert's rationale, see below)	other incorrect comments, incorrect comments without a rationale
Machine learning	ReviewScore (Ryu et al., 2025)	15 graduate students in AI, median of three ratings per premise	lowest premise rating above 3	lowest premise rating below 3	arguments at exactly 3, claims and questions

In the machine-learning part, only arguments (comments that give reasons for a conclusion) are kept, because ReviewScore rates the premises of arguments only. This is why its share of errors is higher.

Human verification

Many comments that the PeerReview Bench experts judged incorrect are errors of judgment rather than factual errors. GPT-5.5 decides from the expert's rationale whether the error is factual (label_origin = expert+llm-filter; all correct rows are expert). Two of the authors independently repeated this decision on the 178 comments that have a rationale. Each saw the comment, the expert rationale and the paper text, but not the GPT-5.5 label or the other annotator's label. The two annotators agree with Cohen's κ = 0.70, and GPT-5.5 agrees with them with κ = 0.76 and 0.77. The released labels are the GPT-5.5 labels.

Machine-learning part (ReviewScore)

ReviewScore states no license, so its comments and papers are not included here. To evaluate on it, download the source data from github.com/Hyun-Ryu/ReivewScore (the repository name is spelled this way) and run scripts/build_reviewscore.py from the PremiseCheck code repository:

git clone https://github.com/Hyun-Ryu/ReivewScore
python scripts/build_reviewscore.py --reviewscore ReivewScore --out data/ReviewFact
# rs: 216 comments, 46 errors, 39 papers  [ok]

The script applies the label rule above and writes the comments and papers in the same format as the other parts (39 papers, 216 comments, 46 errors).

Intended use

ReviewFact is for evaluating and benchmarking systems that detect factual errors in review comments.

Do not use ReviewFact to train models. The licenses of both distributed sources restrict them to evaluation: the ReviewCritique license forbids use for training AI models, and PeerReview Bench is intended for evaluation and not for training.

All parts are test data. The prompts and thresholds of PremiseCheck were developed on ReviewScore and PeerReview Bench, so its results on those parts are not held out.

Licenses and attribution

This dataset is a combination of two derived parts with different licenses (hence license: other). Each part is governed by the license of its source.

• NLP part (data/nlp/): derived from ReviewCritique (Du et al., 2024, github.com/jiangshdd/ReviewCritique), licensed under the Apache NON-AI License 2.0 (see LICENSE_ReviewCritique). Research use only; allowed for evaluation and benchmarking; must not be used to train models. data/nlp/test.jsonl and data/nlp/papers.jsonl are modified files derived from ReviewCritique.jsonl and ReviewCritique_LLM.jsonl. The changes are label mapping, segment selection and new ids.
• Natural-science part (data/science/): derived from PeerReview Bench (Kim et al., 2026, arXiv:2605.20668, github.com/prometheus-eval/cmu-paper-reviewer), licensed under CC-BY-4.0 (see LICENSE_PeerReviewBench.md) and intended for evaluation, not training. data/science/test.jsonl and data/science/papers.jsonl are derived from it; the changes are label mapping, the factual-error filter above, comment selection and new ids.
• Machine-learning part: ReviewScore (Ryu et al., 2025) states no license, so nothing from it is redistributed.
• Overview figure (assets/overview.png): from the PremiseCheck paper.

Known issues

• ReviewCritique segments of fewer than four words (241, all labeled correct, mostly headers such as "Weaknesses:") are removed. Some short header-like segments of four words or more remain (for example "Clarity, Quality, Novelty And Reproducibility:"), all labeled correct.
• Errors are rare (4.8% of NLP comments, 2.5% of natural-science comments, 21% of machine-learning items), and the natural-science and machine-learning parts have only 65 and 46 errors, so confidence intervals on those parts are wide.
• Natural-science error labels depend on an LLM filter over the expert rationale (κ = 0.76 and 0.77 against two human annotators); some disagreement with a human reading is expected.
• Nine natural-science comments have an empty comment string in the source data (prb|37|Human_1|3–5, prb|70|Gemini|1–4, prb|82|GPT|5, prb|82|Gemini|4), all labeled correct. They are kept so that the counts match the paper; you may drop them.
• Paper texts are automatic PDF extractions and contain extraction artifacts (broken small caps, LaTeX fragments, hyphenation).

Citation

If you use ReviewFact, please cite the paper:

@misc{song2026premisecheck,
  title  = {{PremiseCheck}: Efficient Fact-Checking of Human and {AI} Peer Reviews},
  author = {Song, Dingjie and Yan, Zhiling and Zhang, Ziqian and Zhang, Yuxuan and Sun, Lichao},
  year   = {2026},
  note   = {Preprint},
  url    = {https://www.alphaxiv.org/abs/2610.premisecheck-peer-review-fact-checking}
}

Please also cite the source datasets of the parts you use:

@inproceedings{du2024reviewcritique,
  title     = {{LLM}s Assist {NLP} Researchers: Critique Paper (Meta-)Reviewing},
  author    = {Du, Jiangshu and Wang, Yibo and Zhao, Wenting and Deng, Zhongfen and Liu, Shuaiqi and Lou, Renze and Zou, Henry Peng and Narayanan Venkit, Pranav and Zhang, Nan and Srinath, Mukund and Zhang, Haoran Ranran and Gupta, Vipul and Li, Yinghui and Li, Tao and Wang, Fei and Liu, Qin and Liu, Tianlin and Gao, Pengzhi and Xia, Congying and Xing, Chen and Jiayang, Cheng and Wang, Zhaowei and Su, Ying and Shah, Raj Sanjay and Guo, Ruohao and Gu, Jing and Li, Haoran and Wei, Kangda and Wang, Zihao and Cheng, Lu and Ranathunga, Surangika and Fang, Meng and Fu, Jie and Liu, Fei and Huang, Ruihong and Blanco, Eduardo and Cao, Yixin and Zhang, Rui and Yu, Philip S. and Yin, Wenpeng},
  booktitle = {Proceedings of EMNLP 2024},
  pages     = {5081--5099},
  year      = {2024},
  doi       = {10.18653/v1/2024.emnlp-main.292}
}

@article{kim2026peerreviewbench,
  title   = {On the limits and opportunities of {AI} reviewers: Reviewing the reviews of {Nature}-family papers with 45 expert scientists},
  author  = {Kim, Seungone and Yoon, Dongkeun and Gashteovski, Kiril and Suk, Juyoung and Baek, Jinheon and Aggarwal, Pranjal and Wu, Ian and Zaverkin, Viktor and Petkoski, Spase and Schrider, Daniel R. and Dukovski, Ilija and Santini, Francesco and Mitreska, Biljana and Jeong, Yong and Kwon, Kyeongha and Sim, Young Min and Manasova, Dragana and Porto, Arthur and Mojsoska, Biljana and Takamoto, Makoto and Shuntov, Marko and Liu, Ruoqi and Lee, Hyunjoo Jenny and Din{\c{c}}, Niyazi Ulas and Jo, Yehhyun and Han, Sunkyu and Lee, Chungwoo and Li, Huishan and Tsai, Esther H. R. and Simsek, Ergun and Shafi, Khushboo and Chung, Yeonseung and Park, Jihye and Shulevski, Aleksandar and Christiansen, Henrik and Son, Yoosang and Knight, Elly and Montoya, Amanda and Ahn, Jeongyoun and Langkammer, Christian and Moon, Heera and Yoon, Changwon and Stikov, Nikola and Jang, Mooseok and Choi, Edward and Kim, Junhan and Jung, Yeon Sik and Kim, Woo Youn and Kim, Jae Kyoung and Anjum, Ishraq Md and Kim, Hyun Uk and Bridges, Drew and Lawrence, Carolin and Yue, Xiang and Oh, Alice and Asai, Akari and Welleck, Sean and Neubig, Graham},
  journal = {arXiv preprint arXiv:2605.20668},
  year    = {2026}
}

@article{ryu2025reviewscore,
  title   = {{ReviewScore}: Misinformed Peer Review Detection with Large Language Models},
  author  = {Ryu, Hyun and Jang, Doohyuk and Lee, Hyemin S. and Jeong, Joonhyun and Kim, Gyeongman and Cho, Donghyeon and Chu, Gyouk and Hwang, Minyeong and Jang, Hyeongwon and Kim, Changhun and Kim, Haechan and Kim, Jina and Kim, Joowon and Kim, Yoonjeon and Lee, Kwanhyung and Park, Chanjae and Yun, Heecheol and Betz, Gregor and Yang, Eunho},
  journal = {arXiv preprint arXiv:2509.21679},
  year    = {2025}
}