ChicagoHAI/openaireview-perturbation

OpenAIReview Perturbation Benchmark

This benchmark measures how well an AI review system finds errors in scientific papers. We took 72 recent arXiv papers with full LaTeX source and injected known errors into them. A review system reads the perturbed paper, and recall is the share of injected errors that its comments identify.

The benchmark accompanies the OpenAIReview paper and code: https://github.com/ChicagoHAI/OpenAIReview

Contents

• 72 papers from eight arXiv subject classes: cs.CC, cs.LG, econ.EM, hep-ex, math, physics.atm-clus, q-bio.GN and stat.AP.
• 170 rows. Each row is one paper with one error category injected, so a paper appears once per category.
• 2,337 scored errors (2,651 injected in total, see scored below).
• 24 papers form the frontier subset (in_frontier_subset), where the paper also ran GPT-5.5 and Claude Opus 4.7.

Error category	Rows	What changes
Surface	69	A local math edit: a sign or operator, an index or subscript, a numeric parameter, or a computation
Claim	61	A false theoretical claim (a perturbed definition or theorem statement) or a false empirical claim
Experimental	35	A reversed causal claim, a misinterpreted result, or p-hacking
Reasoning	5	Circular reasoning, an invalid implication, an induction error, or a missing case in a proof

Columns

Column	Description
id	<arxiv_id>__<error_type>
arxiv_id, title	Source paper
domain	The arXiv listing the paper was sampled from
primary_category	The paper's primary arXiv category
paper_type	theoretical or empirical, which sets the error types injected
error_type, error_category	Error type of this row (surface, claim_theoretical, statement_empirical, logic, experimental) and its category
in_frontier_subset	Whether the paper is in the 24-paper frontier subset
license, text_included	arXiv license of the source paper, and whether the three text columns are filled
clean_paper	The original paper as LaTeX, flattened by the OpenAIReview parser
perturbed_paper	clean_paper with every edit in perturbations applied
review_input	perturbed_paper cut to its first 13,000 tokens (cl100k_base). This is the text review systems were given
clean_sha256, perturbed_sha256, review_input_sha256	Hashes of the three texts, for checking rebuilt copies
n_perturbations, n_scored	Number of edits, and number that fall inside review_input
perturbations	List of edits, in document order (fields below)

Each item in perturbations has:

Field	Description
perturbation_id	Identifier within the row
error_subtype	Specific error, for example operator_or_sign or causal_reversed
original, perturbed	The text before and after the edit
offset	Character offset of original in clean_paper
why_wrong	Why the edited text is wrong. Scoring compares review comments against this
verifier_quote, verifier_reason	The passage a verifier model used to confirm the error is real, and its reasoning
scored	Whether the edit falls inside review_input and counts toward recall

Getting the text for every row

Only 44 papers (107 rows) carry a license that allows sharing a modified copy (CC BY or CC BY-NC-SA). For the other 28 papers (63 rows), the text columns are empty and the rows keep the edits, offsets and hashes. rebuild.py downloads each paper's source from arXiv, rebuilds all three texts and checks them against the hashes:

pip install openaireview==0.3.0 datasets
python rebuild.py --out rebuilt.jsonl

Running a review system

Give review_input to the review system, then match each comment against the scored edits. The paper counts an error as detected when a comment quotes text overlapping the edit and an LLM judge (Gemini 3 Flash) rates the comment's explanation at least 4 out of 5 against why_wrong. The scoring code is benchmarks/perturbation/score.py in the GitHub repository.

How the errors were made

Gemini 3 Flash proposed edits for candidate spans in each paper, and Claude Sonnet 4.6 kept only edits it judged to be real and substantive errors. Papers were then cut to 13,000 tokens, and rows with fewer than 5 errors left in that range were removed. We sampled 74 papers, and 2 lost every row to that filter, which leaves 72. One edit replaced \right with a carriage return followed by ight, from a JSON escaping error during generation. The benchmark read papers in text mode, so the carriage return became a newline, and the data here matches what review systems saw.

License

Each row is a derivative of one arXiv paper and keeps that paper's license, given in the license column. Rows from CC BY-NC-SA papers may be used only for noncommercial purposes and shared under the same license.