43,013 ICLR 2027 submissions. In public.

Red-flag Research

Live, public screening for automated-research attacks on peer review

A research project of OpenLAIR, by Ziqian Zhang.

--d --h --m --s

until Stage 1 starts

43,013ICLR 2027 submissions

Red-flag Research is a research project from OpenLAIR that screens conference submissions for automated-research attacks, starting with paper laundering in ICLR 2027. It checks each submission against earlier public papers and flags close matches that lack clear credit, so a person can compare the two side by side. A flag is a lead, not a verdict.

Public: benchmark, evaluation tools and saved baseline outputs.

9public stages up to 5,000 papers per stage —papers checked

What is a red flag?

A red flag asks a person to look at two papers side by side. Here is what we look for, and what a flag does and does not mean.

What we study

Automated-research attacks

AI tools can now draft, rewrite and package papers at scale. Some of that output takes credit for other people's work. We call this kind of misuse an automated-research attack.

The first kind we screen

Paper laundering

Earlier work rewritten and repackaged so it reads as new: the same results, tables or contributions in different words, without a clear citation of the source.

What a flag means

Source review recommended

When passages, numbers or tables closely match an earlier public paper without clear attribution, the paper gets a red flag. It asks a person to compare the two. It is not a finding of misconduct.

Not a red flag: the paper's own earlier preprint, shared authors, a cited dataset or baseline, or reuse that is clearly attributed. The system can miss these, which is why every flag goes to a person.

What it costs

On our benchmark, doing every step with a large model costs 1.5–18× our pipeline, depending on the model and how it batches.

$0.047

Our pipelineModel API fees per paper, mean over 19 benchmark papers.

The same pipeline, with a large model doing every model step
Large model for every step	Per paper	vs our $0.047	Batched, 8 per call	Rangemultiple of ours, log scale1× ours5×10×20×
Qwen3.8-flash low	$0.26	5.6×	$0.071 · 1.5×	1× ours20×
Qwen3.8-flash xhigh	$0.61	13×	$0.12 · 2.6×	1× ours20×
GPT-6 luna max	$0.86	18.5×	$0.15 · 3.3×	1× ours20×

What goes into these numbers

$0.047 is the mean model API cost per paper measured on 19 benchmark papers. It does not include OCR or search.
The large-model rows run our own pipeline, step for step, with the large model in place of our models. Each paper has about 138 candidate sources. We screen their abstracts with a small, cheap model and send only 3 to full-text comparison. The large-model figures have it read every candidate.
The full-text pair cost is measured: $0.010–0.013 per pair for the large models, $0.005 for ours.
Costs were paid in CNY and are shown in US dollars at ¥7.0 = $1.

Nine stages, back to back

All 43,013 submissions in nine stages. Times are US Eastern; those marked est. are estimates and move as stages finish. Amber flags mark ICLR 2027 dates for reference.

How many red flags?

Our estimate of how many of the 43,013 submissions will end with Source review recommended, and how the running count compares as the stages go by.

Your prediction

How many submissions will end with a red flag?

Pick a range. You can change or take back your pick until the poll closes. No account; your browser keeps a random id so you can change your pick.

Our estimate before the run

—

043,013

Estimated rangeProjection from the running count

Running count

—

Counted from the public live data. A flag can change after human review or a correction, so the count can go down as well as up.

Notes from the run

Newest first. Numbers come from the run records; individual papers appear only on the live page.

Watch it as it runs

The live page follows the running sessions one paper at a time, from the first claim to the final label.

Live page opens when Stage 1 starts

For each submission you can see:

The claims the pipeline extracts from the submission.
The earlier public papers it finds and ranks.
The closest passages, side by side, with model scores.
The result label, with a link to report a mistake.

Open the live page →

A flag is a lead

What a label means, what it does not mean, and how to get a result corrected.

A lead for review, not a verdict

Red-flag Research is a research project that screens conference submissions automatically. A Source review recommended label means the system found passages, numbers or tables that closely match an earlier public paper without a clear attribution, so a person should compare the two papers side by side.

It is not a finding of plagiarism, misconduct or bad faith. It has not been reviewed by the conference, the authors or any integrity committee. The system uses language models and can be wrong: shared authors, an earlier version of the same work, a cited dataset or a common baseline can all explain matching text, and the system may not see them. A paper without a label has not been cleared either.

Benchmark cases are anonymized before sharing.

Result labels

Source review recommendedClose matches to an earlier public paper without clear attribution. A person should compare the two.
No source findingNothing found in the sources the system checked. This does not clear the paper.
InconclusiveNot enough evidence for either label.
Same paper (preprint)The closest match is this paper's own earlier preprint.

Think a result is wrong? If you are an author and believe a result about your paper is wrong, tell us through the feedback form or at red.flag.autoresearch@gmail.com. We will correct it right away, and we take the result off the live page and the public data while we check.

Report a wrong result

The benchmark

An anonymized benchmark for this task is available on GitHub and as a Hugging Face dataset, not on this site. Cases are anonymized before sharing.

huggingface.co/datasets/red-flag-research/iclr2027-source-review-benchmark

The same benchmark as a Hugging Face dataset, CC BY 4.0.

Open on Hugging Face →

Red-flag Research Source-Review Benchmark (ICLR 2027 diagnostic set)

What is in it: 20 anonymized ICLR 2027 target papers, 34 target–source pairs, a human label for every target, evidence locations, evaluation tools and saved baseline outputs.
Labels: 16 No source finding, 2 Same paper (preprint), 2 Source review recommended. Labels apply to target papers, not to each pair, and these cases are separate from the 43,013-submission public screening.
Limitations: One annotator, two positives, and model-influenced selection. The set does not measure how common source reuse is or how accurate a detector is in general.
Version and licenses: Version 0.3.0. Annotations and labels CC BY 4.0; evaluation tools Apache-2.0; source papers keep their own licenses. Labels are leads for human review, not verdicts.

What comes next

Features we plan to add to make the screening more useful and easier to check. Plans can change; this list is updated as they do.

Questions

Short answers about what the project does, what is public, and how to get a result corrected.

What is Red-flag Research?

Red-flag Research is a research project from OpenLAIR that screens conference submissions for automated-research attacks, starting with paper laundering in ICLR 2027. It was started by Ziqian Zhang at OpenLAIR (Lehigh AI Research Lab) in October 2026.

What is paper laundering?

Paper laundering is rewriting earlier public work into a submission that looks new, without credit to the original. Language models make the rewriting cheap, and text-similarity checkers can miss it because the wording changes while claims, numbers and tables stay close to the original.

Is Red-flag Research a plagiarism checker?

Not in the usual sense. Similarity checkers such as iThenticate look for copied wording. Red-flag Research compares claims, passages, numbers and tables with earlier public papers even after rewording, and it never decides whether anything is plagiarism. A person does.

What does “Source review recommended” mean?

It means the system found passages, numbers or tables that closely match an earlier public paper without a clear attribution, so a person should compare the two papers. It is not a finding of misconduct and has not been reviewed by the conference, the authors or any integrity committee.

Which papers does it screen, and when?

The first application screens all 43,013 ICLR 2027 submissions in nine public stages of up to 5,000 papers. Stage 1 starts on October 16, 2026 at 9 PM US Eastern, and the run aims to finish before ICLR 2027 decisions on December 16, 2026.

What is publicly available?

The benchmark (20 anonymized ICLR 2027 cases with human labels), the evaluation tools and the saved baseline outputs are public at github.com/OpenLAIR/red-flag-research and on Hugging Face (red-flag-research/iclr2027-source-review-benchmark).

Is Red-flag Research affiliated with ICLR?

No. It is an independent research project from OpenLAIR. It uses submission metadata that is public on OpenReview and does not host papers.

How do I get a result corrected?

Use the feedback form on the site or e-mail red.flag.autoresearch@gmail.com. The result comes off the live page and the public data while we check, and we correct it right away if it is wrong.

Help us get it right

Corrections, bug reports, benchmark data and ideas all go through one short form. It opens your own email app with a tagged message, so nothing is posted in public.

First: a result about a paper looks wrong

Tell us which paper and what is wrong. You do not need an account, and the message stays private. We take the result off the live page and the public data while we check, and correct it right away if it is wrong.

Or email red.flag.autoresearch@gmail.com directly.