AI-powered feedback trusted by world-class researchers
Refine devotes hours of compute to help you find and fix the issues that matter most to readers and reviewers.
Empirical specification description
"We use the inverse of the propensity scores as weights."
For ATT, control units should be weighted by ωᵢ / (1 - ωᵢ), not 1 / ωᵢ, where ωᵢ is the propensity score of unit i. Suggestions: if the nonstandard weights are appropriate in this context, explain them when the weights are discussed.
Address case when x is negative in Proposition 2.
"Lemma 2. The optimal level y(x) satisfies..."
The proof of the lemma implicitly assumes the argument x is nonnegative, but the maintained assumptions do not require this. Suggestions: either add an argument to deal with the negative case, or make an assumption to rule out negative x.
Trusted by researchers at top institutions
Designed for researchers who demand rigor
Depth
Uncover errors before they cause headaches
Refine helps you check your paper before submission and avoid the kinds of avoidable mistakes that cause stress and lost time later.
"spotted a number of errors"
I have found Refine to be very useful. It has spotted a number of errors, both obvious ones, and some that are fairly subtle, and it does a great job of checking for consistency between what is said in various parts of a paper.
Drew Fudenberg
Paul A. Samuelson Professor of Economics
MIT
Precision
Tap into a level of analysis unmatched by any general-purpose AI
Refine devotes hours of parallelized compute to deeply understand your arguments. It produces targeted feedback — no slop, no generic suggestions, just the comments that would make a reviewer reconsider.
"Refine is amazing!"
Refine is amazing! I've used it on projects ranging from the least to the most technical parts of philosophy, and in every case it's found errors I had missed, from small issues at the sentence or paragraph level to errors in keeping track of cases in longer proofs, all the way to mistakes in big arguments or how I cited the literature. The gain over publicly available tools is very significant. I plan to run every paper I write through it at least once before submitting.
Harvey Lederman
Professor of Philosophy
UT Austin
Speed
Reviewer-grade results in minutes, not months
Most papers return in 20–40 minutes. You get reviewer-grade insights long before a peer-review cycle would surface them, helping you iterate fast.
"would have taken an expert many hours to complete"
The depth of the reading performed by Refine was impressive, and resulted in the identification of subtle issues. This would have taken an expert many hours to complete, even when working with ChatGPT Pro.
Omer Tamuz
Professor of Economics and Mathematics
Caltech
Privacy, by design
Your research stays yours—private, secure, and never used to train AI
Your content is completely secure
We protect your content with strong security controls and encrypted storage, following modern best practices. Learn more in our Privacy Policy.
Never used for training
Your papers are yours. Period. Refine's model providers operate under zero-retention contracts — nothing you upload ever becomes training data.
SOC 2 + ISO 27001
Pursuing certification in both. Our full security posture — subprocessors, encryption, incident response — is public on our Trust Center.
How a Refine review is produced
Every review leverages massive parallel compute to deliver feedback worth your attention.
2+
hours
Of parallel compute per review
Hundreds
of frontier LLM calls
Stress-test every line of your document
~30
minutes
Average end-to-end turnaround
Questions