Guide
Review my paper before submission: what can be checked, and by whom
Yes, a tool can review a draft before you send it, but only the third of the work that comes from reading it: whether the conclusion follows from the evidence, whether the statistics hold, whether the contribution is new against the literature, whether anyone could rerun it. The other two thirds are an administrative layer only you can do because none of it is in the file, and a mechanical layer a compiler and a link checker settle in minutes, and treating all three as one job is how the last week before a deadline gets spent on the wrong one.
Published 19 August 2026
The short version
- Do the administrative layer first. It depends on other people and on systems that close early.
- Have the argument read second. While the draft is still soft enough to change: the objections that come back are days of work, not minutes.
- Do the mechanical layer last, on the exact file you upload. Every one of those checks is undone by the next edit.
- Only the middle step is something we sell. The two around it are yours whatever you buy.
The division of labour
Three layers, and only one of them can be bought
The useful question is never which tool. It is which of these three you are actually short of this week, because the answer picks a different person every time.
| Layer | What it covers | Who does it |
|---|---|---|
| Eligibility and consent | Co-author registration, submission-system profiles, conflicts, submission caps, dual submission, data and subject permissions, the anonymity of your own repository. | You, and nobody else. None of it is in the manuscript, so nothing that reads the manuscript can see it. |
| Mechanical conformance | Page count against the current call, an unmodified template, required statements, a bibliography that compiles clean, links that still resolve, identity left in the PDF properties. | A compiler, a link checker and ten minutes. No judgement is involved, which is why it is the cheapest layer and the last one to do. |
| The argument | Whether the conclusion follows from the evidence, whether the contribution is what the paper says it is, whether the statistics survive, whether it is new, whether anyone could rerun it. | Someone who reads it with the venue in mind: an advisor, a colleague outside your group, or this. |
The layer people skip is the one with no deadline of its own
Registration closes on a date and the template either compiles or it does not, so both of those get done. Nothing forces you to find out, before you send it, that your conclusion is wider than your experiment supports. That check has no external deadline, which is exactly why it is the one that slips, and it is also the only one whose findings need a week to act on.
A perfectly conformant paper still gets rejected on the argument
Formatting compliance is a floor, not a signal. Once the file is inside the limit and inside the template it stops earning anything, and the rest of the decision is made on grounds a formatter cannot inspect. Spending the final week on the floor is the most common way to arrive at the deadline with a submission that is correct and unconvincing.
Layer one
The checks that are not in the file
Every item here lives in a submission system, an inbox or a legal document. A reader of the PDF, human or otherwise, can see at most that you claimed it.
| What to check | Why nothing that reads the paper can see it |
|---|---|
| Every co-author has an account, and it is complete | Profiles and reviewer registration live under accounts nobody but their owner can open. A co-author who has not finished registering on the last evening is a submission that does not go in, and it is the failure mode nobody rehearses. |
| Conflicts of interest are declared on each profile | Conflicts come from employment and collaboration history that exists nowhere in the manuscript. Declaring them badly hands your paper to somebody who should not be reading it, and fixing that after assignment is somebody else's decision. |
| You are inside the venue's submission cap | Caps count per person across the whole venue, so the number lives in the submission system rather than in any one paper. Some venues also tie reviewing duty to how many papers you submitted. |
| The work is not under review anywhere else | Dual submission is a fact about your other papers and their timelines. It is a policy violation rather than a formatting slip, and the consequences outlive the paper that triggered them. |
| You have the permissions the paper implies | Data licences, redistribution rights, consent and ethics approval are documents outside the PDF. A reviewer reads your statement about them; only you can read the approval itself. |
| Your repository is anonymous, if the review is double-blind | A commit history, an account name and an organisation page sit in a different system entirely, and this is where anonymity leaks most often. At a single-blind or open venue the same link is expected instead, so establish which kind you are submitting to before you strip anything. |
| The author list is final | Many venues freeze the author list at the deadline and reopen it for nothing, camera-ready included. It is a five-minute conversation that becomes impossible one minute later. |
Requirements differ between venues and between editions of the same venue, so this table names the check and never the number or the date. The current call for papers is the only authority on both. The fuller list of failures, ordered by how far into the process each one kills you, is the pre-submission checklist.
Layer two
What a script settles, and what we do not pretend to do for you
This layer is mechanical, cheap and genuinely worth doing. It is also the layer where we are honest about not being the answer.
| The check | How to settle it in minutes |
|---|---|
| Length against the current call | Count what the call counts, on the final build. References, appendices and required statements are treated differently by different venues and the rules move between editions. |
| An unmodified template | Build with the style file taken from the venue this cycle, and remove every margin and spacing hack before the last compile. Shrinking anything to fit is read as gaming rather than as an accident. |
| Required statements present | Reproducibility, ethics, limitations, a checklist: whichever the current call names, and often outside the page limit. They are short, they are mandatory, and they are the ones people leave to the last hour. |
| A bibliography that compiles clean | Read the compile log rather than the PDF. Every citation resolves, every entry is cited, no undefined reference survives into the build you upload. |
| Links that still resolve | Fetch every URL in the paper once, including the anonymised mirror. A repository link that fails during the review window costs more than the missing repository would have. |
| No identity left in the file | Open the compiled PDF properties and read the author field your editor wrote without asking, then search the source for your surname, your institution and your funder. |
| Figures readable without colour | Print one page in greyscale. It takes a minute and catches a figure that was unreadable to a share of your reviewers. |
We built checkers for this layer and they are not in the paid path
Format, structure, spelling and prompt-injection agents exist in the codebase, start with the system and receive no traffic today, because the orchestrator dispatches only to the six specialists that reach a verdict. We would rather write that down than let the sentence "it checks your formatting" sit here unearned. If you upload a PDF the point is moot anyway: the LaTeX source is gone, and page geometry and section markup cannot be recovered from a rendering of them.
The citation checker is switched off on purpose
It was built, it worked on LaTeX sources, and on extracted PDF text it produced false hits often enough that reporting them would have cost a reader more time than the tool saved. A tool that flags a tensor shape as a missing citation trains you to ignore its output, so it stays off until the extraction is good enough. Bibliography completeness stays a compile-log job, which is where it is reliable.
Layer three
The part that needs someone to read it, and what we do there
Six specialists read the same draft with a different question each, and an orchestrator turns the six answers plus the guidelines of the venue you picked into one call.
| Specialist | The question it asks of your draft |
|---|---|
| Soundness | Does each claim have evidence behind it, and does the conclusion follow from what was actually shown rather than from what the introduction promised? |
| Significance | Is the contribution the kind of thing the paper says it is, or is an incremental benchmark delta being presented as a method? |
| Statistical | Do the numbers survive the way they were produced: seeds, variance, the baselines chosen, and whether the comparison being drawn is the one the table supports? |
| Coherence | Does the paper read as one argument, or do the abstract, the method and the conclusion describe three papers with different scopes? |
| Novelty | Is this new against the literature as it stands today, searched live across Semantic Scholar, arXiv and OpenAlex rather than recalled from training data? |
| Reproducibility | Could a stranger rerun it: code, data, hyperparameters, compute, and do the artefact links actually resolve? |
An orchestrator, not an average
The six results go to a reviewer agent that loads the guidelines of the venue you selected from the database and decides against them, because the same paper is a different proposition at a venue that rewards ambition and at one that rewards completeness. It returns accept or reject. Binary on purpose: a number between one and ten reads as a prediction of a review score, and it is not one.
What the report gives you, item by item
Per specialist a score, a summary and the objections it raised, and each objection carries a severity, a category, what is wrong and a suggested fix. It does not contain a rewrite of your paper, a predicted review score or a guess at your reviewers, because none of those are ours to give.
It was scored against decisions that already happened
697 real accept and reject decisions across ICLR 2017, NeurIPS 2023 and ICLR 2025: AUC 0.91 and 96% accuracy on the 297 ICLR 2025 papers, with one false accept among 149 rejects. On that set it is far likelier to warn you about a paper that was accepted than to wave through one that was not.
A specialist that comes back useless fails the run
If an agent times out, errors or returns nothing usable after a retry, the review is marked failed and the credits go back automatically. A verdict assembled from four of six opinions and presented as six would be the one failure mode you could not detect from the report, so it does not happen.
Sequencing
With a week left, the order decides the outcome
Every item below is worth doing. Done in the wrong order, half of them are wasted, because the cheap checks are only true about a file that has stopped changing.
| When | What | Why then and not later |
|---|---|---|
| First, today | Layer one: accounts, profiles, conflicts, caps, permissions, the author list. | All of it depends on other people or on a system that closes early, and none of it can be compressed on the last evening. It is also the only layer where a single missing item ends the submission outright. |
| Second, while the draft is still soft | Layer three: have the argument read, against the venue you are actually aiming at. | The objections that come back are days of work, not minutes. Getting them on the final morning means getting them too late to use, which converts the most useful check into an expensive way to feel bad. |
| Then, for as long as you have | Fix in severity order, hardest objection first. | An objection that says the conclusion outruns the evidence is a rewrite of a section or an extra experiment. One about a caption is ten minutes. Doing the ten-minute ones first feels like progress and buys nothing. |
| Only if the argument changed | A second read. | A second pass over the same claims returns the same objections. It costs another 60 credits, so it is worth it when a section was rewritten or an experiment was added, and worth nothing when the diff is typography. |
| Last, on the exact file you upload | Layer two: length, template, statements, compile log, links, PDF properties, greyscale. | Every one of these is undone by the next edit. Running them before the final build proves something about a file that no longer exists, which is the most common way a paper goes over length after being checked. |
| The one you cannot recover | Abstract registration. | At the large ML venues it closes before the paper deadline, and a paper with no registered abstract has nothing to attach to. Put that date in the calendar as the real deadline and treat the paper deadline as the second one. |
The boundary
What none of this does, including ours
Stated here rather than at the bottom of a sales page, because a reader who finds the limits later stops believing the rest.
It does not tell you whether your paper will be accepted
The verdict is a judgement about the draft, benchmarked against decisions that were already made by other committees on other papers. Your decision depends on the reviewers you draw, the pool you land in and the year, and the same paper sent twice can go both ways without anybody being at fault.
It does not replace the reader who knows your subfield
An advisor or a colleague outside your group catches the thing nobody else can: that the framing is wrong for this community, that the obvious baseline is missing because it is obvious only to five people, that the related work leaves out the group who will be reviewing you. Buy the machine read to save that person's attention for the questions only they can answer, not to avoid asking them.
It does not write the paper, and it should not
The output is a list of objections with severities and suggested fixes. What to do about each of them is authorship, and venues increasingly ask you to disclose what was generated, so keep the boundary where it is easy to describe.
It does not talk to the venue on your behalf
Nothing here submits, registers, withdraws or corresponds with a chair. Every action inside the submission system stays with the author, which is also why layer one can never be delegated.
The longer version of this argument, with the objections stated by the people who raise them, is what AI peer review cannot do. For the reasons papers fail once they are inside, see why papers get rejected.
Questions
Asked the way people ask it
Short answers, with the concession in the first sentence rather than the last.
Is there a tool that reviews my paper before I submit it?
Yes, for the part of the work that comes from reading the paper. A tool can tell you whether your conclusion outruns your evidence, whether the statistics survive the way they were produced, whether the contribution is new against the current literature and whether a stranger could rerun it. It cannot register your co-authors, confirm your ethics approval or anonymise your repository, because none of that is in the file. Ours reads the draft with six specialists, applies the guidelines of the venue you pick and returns accept or reject with the objections behind it, in about four minutes, for 60 credits, which is €9.
Can AI review my paper before submission?
It can read it and object to it, which is the useful half. What it cannot do is the administrative layer, which lives in the submission system rather than the manuscript, and it should not be doing the writing. Venue policies on language models are mostly written about reviewers handling other people's submissions and about disclosing generated content, and reading a report about your own draft is a different act, but the current call for papers is the only authority on your venue and the policies change every edition.
What should I check before submitting to a conference?
In this order. First the things only you can check: every co-author registered, conflicts declared, the submission cap, dual submission, data and ethics permissions, an anonymous repository if the review is double-blind, and the author list you want frozen. Second the argument, while the draft can still change: does the conclusion follow, is the contribution what you say it is, do the statistics hold, is it new, can it be rerun. Last, on the exact file you upload, the mechanical layer: length, unmodified template, required statements, a clean compile log, working links, no identity in the PDF properties. Abstract registration usually closes before the paper deadline, so treat that date as the real one.
Will an AI review tell me if my paper will be accepted?
No, and treat anything that promises it carefully. Our verdict is accept or reject and it was scored against 697 decisions that had already been made: AUC 0.91 and 96% accuracy on the 297 ICLR 2025 papers, with one false accept among 149 rejects. That is a measurement against past committees, not a forecast of yours, since the same paper sent to two committees can go both ways and nothing in a report changes the reviewers you draw. The usable output is the list of objections, answered while the paper is still yours.
How long does a pre-submission review take, and what does it cost?
About four minutes for a single pass over one paper, and 60 credits, which is €9. Matching a paper against the venues that publish its subject is free and needs no account. A second pass is worth paying for when a section was rewritten or an experiment was added, and worth nothing when only the typography changed.
What is actually in the report?
A verdict, accept or reject, decided against the guidelines of the venue you selected. Then, per specialist, a score, a summary and the objections it raised, each carrying a severity, a category, what is wrong and a suggested fix. It does not contain a rewrite of your paper, a predicted review score or a guess at who will review you. If a specialist fails or returns nothing usable, the run is marked failed and the credits are returned rather than a partial verdict being presented as a whole one.
The objections are cheaper now than in the reviews
Layers one and two are yours and cost nothing but attention. The middle one is a draft read against your venue's own guidelines, returned while the paper is still in your hands.
phdflow reads a finished draft against the guidelines of the venue you are aiming at and returns an accept or reject call with the objections behind it. Scored against 697 real decisions: AUC 0.91 and 96% accuracy on 297 ICLR 2025 papers, one false accept in 149 rejects. The full measurement. Read a finished review before you pay for one. Your first paper gets a free preview; the full report is 60 credits, €9. No account needed. Pricing.