FactReview
Built onFactReviewXu et al., 2026arXiv:2604.04074 ↗
Peer review, with evidence.

Every claim, traced to its source.

Upload an ML paper. We extract every claim, locate it in the literature, verify each citation, and return a structured report you can audit line by line.

Encrypted in transit & at restNever used to train AIFree for academics
See an example review →no signup needed

Run a review

factreview.ai/submit

Drop a PDF
Drop a PDF here, or click to choose
PDF · up to 30 MB
or
Paste an arXiv URL
We'll fetch the latest version automatically.
Your email
We'll mail you a magic link when the review is ready.
Average runtime ≈ 25 min · You'll be emailed when done
01 / Pipeline

How it works

PDF
Upload paper
arXiv · ≤30 MB
01
Parse
Extract sections, tables, citations and figure regions from the PDF.
02
Extract & locate
Pull every empirical, theoretical and reproducibility claim. Pin each to a section.
03
Verify
Cross-check against literature, RefChecker and (optionally) executed code.
04
Report
Each claim gets a 5-label verdict, evidence list and a one-glance teaser figure.
JSON
Structured review
queryable · exportable
02 / Differentiation

What you get that others don't

FactReview
this product
Stanford AgenticManusightsReviewer3
Citation verification, no capyesno~partialno
Literature positioning + novelty verdictyes~partialnono
5-label claim ledger w/ evidence linksyesno~partial~partial
Teaser figure (poster-style overview)yesnonono
Bilingual output (EN / 中文)yesnonono
03 / Trust

What we promise

01Your paper is encrypted in transit and at rest. Always.
02We will never use your paper to train a model. Ever.
03Your review URL is a 32-byte secret — unguessable, never indexed.
04 / FAQ

Frequent questions

Is FactReview the same as the FactReview paper?

We're built on top of it. The pipeline is from Xu et al. (2026); we wrap it in a web product. Every review is stamped with the backend version that produced it.

Can I use this for an OpenReview submission?

Yes — but check your venue's disclosure rules. We surface that reminder in the workspace and in every email footer.

Why not give me a single accept/reject score?

Because that's the part you should decide. We collect evidence; you read it. AI in peer review is most useful as an evidence collector, not a judge.

What languages do you support?

Input must be an English ML paper. Output ships in EN and 中文 — evidence quotes always stay in the original English so they remain traceable.