AI peer review before submission
Know the verdict before the reviewers do
Upload your draft and pick the venue. Six specialist agents read it, and a reviewer weighs them against that venue's criteria: accept or reject, with every issue located, in about four minutes.
For ICLR, NeurIPS, ICML, CVPR, ACL ARR, AAAI, TMLR and more. Your first paper is read at no charge, and the report opens for €24.99.
New review
Submit a paper
Your draft is never published. The report opens only through its own private link. What happens to it
- AUC 0.91
- on 297 ICLR 2025 papers, 95.96% accuracy
- About 4 minutes
- from upload to verdict, and you watch each agent finish
- Private draft
- never published, read through a private link
What the report looks like
A real run on arXiv 2502.04180, target venue ICLR 2025
Verdict
Reject
Overall score
0.59
The proposed 'textual gradient' (Eq. 12) is a heuristic, not a gradient, yet it is presented as backpropagation for black-box operators.
Suggested fix: Clarify that this is a heuristic search method rather than gradient-based learning.
Results are reported as point estimates without confidence intervals or variance across seeds, despite the stochastic controller network (Eq. 7).
Suggested fix: Include standard deviations over at least 3-5 independent training runs.
Two of the fourteen issues this run found. It was made before the reproducibility agent joined, so it has five scores, not six. Read a real, unedited review
How a review runs
- 1
Upload and pick the venue
A PDF of the draft, the venue you are aiming at, and the address the report link goes to.
- 2
Six specialists read it
Soundness, significance, statistics, coherence, novelty and reproducibility, each scored on its own. The novelty agent searches the literature for prior work.
- 3
A reviewer decides
It weighs the six against that venue's own criteria and returns accept or reject, with every issue located in the paper.
Your draft stays private
- It is never published, and the report opens only through its own private link.
- The PDF is parsed on our own servers. The text is read by Google's Gemini API to produce the review.
- The novelty check searches Semantic Scholar and arXiv for related work.
That is the whole list, as our privacy policy sets out. To have a review deleted, write to [email protected].
Checked against real decisions
We ran it on 697 papers whose accept or reject decision is public, from three venues, and compared its call with the real one.
How accurate is AI peer review, measured The confusion matrix
Questions before you upload
What happens to my unpublished draft?
It is never published, and the report opens only through its own private link. The PDF is parsed on our own servers, the extracted text is sent to Google's Gemini API to run the review, and the novelty check searches Semantic Scholar and arXiv for related work. That is the whole list, and the privacy policy sets it out. To have a review deleted, write to [email protected].
What do I get back?
An accept or reject call for the venue you picked, an overall score, a score from each of the six specialists (soundness, significance, statistics, coherence, novelty and reproducibility), and every issue they found with the section it sits in and a suggested fix.
What does it cost?
The reviewers read your first paper at no charge. Opening the report they write is €24.99, or 60 credits, and nothing of the result is shown before that. If your first review is not worth what you paid for it, write to us within 14 days and we refund it. One per account, no form to fill in.
Do I need an account?
No. You give an email address and the link to your report is sent to it. No password, and the report also opens in the same tab when it is ready.
Which venues can I choose?
The machine learning and AI venues people submit to most, including ICLR, NeurIPS, ICML, AAAI, IJCAI, CVPR, ICCV, ECCV, ACL Rolling Review, TMLR, COLM, AISTATS and UAI, plus a few outside computer science. The reviewer weighs the specialists against the criteria of the venue you pick.
How accurate is it?
It was scored against 697 real accept and reject decisions from three venues. On the 297 ICLR 2025 papers it reached AUC 0.91 and 95.96% accuracy; across all three venues, 91.68% accuracy and AUC 0.87. The confusion matrix is on the architecture page.
Will using this get my paper desk-rejected?
No. Venue policies target two things: reviewers running confidential submissions through an LLM, and authors passing off generated text as their own writing. This is neither. You run it on your own manuscript, before submission, for feedback. Disclosure rules are your venue's and they change every cycle, so read them: we are not going to claim blanket compliance on your behalf.
Find out before the venue tells you
A missing baseline costs a weekend before the deadline and a whole cycle after it. The first read is free and takes about four minutes.
The reviewers read your first paper at no charge. Opening the report they write is €24.99, or 60 credits, and nothing of the result is shown before that.