Guide
The pre-submission checklist
Ordered by how far into the process each failure kills you. The first group ends your submission before a reviewer is assigned, and returns no feedback at all, which makes it the only part of this page that is genuinely urgent.
Published 11 August 2026
Draft finished, not sent. This is the week where every failure below is still free to fix. If the paper has already gone out, the page for that week is how to write a rebuttal
If you have one hour
- Do the first table and nothing else. The rest can wait for an hour you do not have.
- Everything below it costs you a score. Everything in it costs you the whole cycle.
- Six months is a long time to pay for a filename. That is the entire argument for doing this part first.
Before anyone reads it
The failures that end a submission without a review
These are handled administratively. There is no discussion phase, no rebuttal and no reviewer to persuade, and at most venues no appeal worth the name.
Three of these apply only if the review is double-blind
Plenty of venues are not. Most journals and a good number of conferences run single-blind or open review, and there your name, your affiliation and a link to your repository belong on the manuscript and are often required. The anonymity rows below are about the other kind, where the venue removes identity on purpose and checks it before a reviewer is assigned. ICLR, NeurIPS and the large vision and language conferences work this way. Establish which kind you are submitting to before you use this table, because the advice inverts completely between them.
| The failure | Why it ends there | The check that catches it |
|---|---|---|
| The abstract was never registered | Both major venues close abstract registration before the paper deadline, and a paper with no registered abstract has nothing to attach to. | Put the abstract deadline in your calendar as the real deadline. It is the one that cannot be recovered. |
| Double-blind only: your identity is in the file | Where a venue runs double-blind, this is checked before assignment and handled administratively, so it makes no difference that the leak was an accident. | Search the source for your surname, your institution and your funder. Open the compiled PDF properties and read the author field your editor put there without asking. |
| Double-blind only: a link identifies you | A repository URL carrying your username, or a personal site, undoes the anonymity the venue is enforcing. At a single-blind or open venue the same link is expected and often required. | Where review is double-blind, route external links through an anonymised mirror and hold the acknowledgements and funding lines back until the camera-ready. |
| Double-blind only: self-citation in the first person | "In our previous work" names the authors without naming them, and reviewers notice immediately. Outside double-blind it is simply normal writing. | Search for "our previous", "we showed in" and "our earlier". Cite yourself in the third person like anybody else. |
| The paper is over length | Length is checked mechanically against the current limit, before assignment. | Count what the call counts. Read the current call for what is inside the limit and what is not, because appendices, references and required statements are treated differently and the rules move. |
| The template was altered | Margins, spacing and font size are checked, and shrinking any of them to fit is treated as gaming rather than as an accident. | Compile with the unmodified style file from the venue and remove every spacing hack before the final build. |
| A required statement is missing | NeurIPS has asked for a paper checklist and a limitations section; ICLR has asked for reproducibility and ethics statements, outside the page limit. | Open the current call and write these before the last day. They are short, they are required, and they are the ones people skip. |
| The same work is under review elsewhere | Concurrent submission is a policy violation rather than a formatting slip, and the consequences run past this paper. | Check the venue's dual submission rule against every workshop and archival preprint the work has been through. |
Compiled from the published author instructions of ICLR and NeurIPS, both of which run double-blind. Requirements differ between venues and between editions of the same venue, and this page carries no page limits, font sizes or dates on purpose, because those change every edition and a checklist confidently wrong about one causes the exact failure it exists to prevent. Take every number from the current call for papers.
Still fixable
The things that cost you a score rather than the cycle
These are the objections reviewers write most often, in the order that gives you the most back for the time you have left.
Audit the abstract against the figures
Read your abstract as an adversary and mark every noun that promises a scope: generalisation, robustness, efficiency, any domain named in the plural. Check each one against a figure or a table. Anything with nothing behind it either gets evidence or gets narrowed. This single pass removes the most common rejection sentence there is, and it costs no experiments.
Make the contribution one sentence a stranger could repeat
Not a list of three bullets describing what you built. One sentence naming the thing that was not possible before. If you cannot write it, the reviewer who could not find it was reading accurately, and no amount of related work will cover for it.
Add the baseline you are hoping nobody asks about
You already know which one it is. Either the row goes in the table, or one sentence in the paper says why the comparison does not apply. An explained absence costs nothing and an unexplained one costs the reader's trust in every other row.
One ablation per component you named
A system with several new parts and a single end number invites the reading that the gain came from something incidental. The table that separates them is usually a day of compute and it answers an objection you cannot answer later.
Write the limitations section as though you meant it
A limitations section listing only things that are not really limitations reads worse than none, and at venues that require it a reviewer will say so. Naming the real boundary of the work is also the cheapest way to disarm the reviewer who was going to name it for you.
Say where the code, the data and the settings are
Hyperparameters, hardware, wall-clock, seeds, and an anonymised repository if the venue allows one. Reproducibility questions are the easiest points on the board and the most commonly left on the table.
The first five minutes
A reviewer forms the verdict early and spends the rest confirming it
Nobody reads a submission front to back on the first pass. They read the title, the abstract, the first figure, the main table and the conclusion, and by then they know what they think. These five are worth a disproportionate share of your remaining time.
Figure 1 has to carry the idea alone
It will be read before the method section and often instead of it. If it needs section 3 to make sense, it is a diagram of your implementation rather than of your idea, and the reader who could have been convinced in ten seconds is now reading prose instead.
Every caption stands on its own
A caption that says "results on CIFAR" wastes the one line guaranteed to be read. Say what the reader is looking at and what it shows, in a sentence that works without the surrounding text.
The main table needs its answer marked
If the reviewer has to work out which column matters, some of them will decide wrongly. Bold the result, order the rows so the comparison is adjacent, and keep the units and the direction of better visible.
Print it in greyscale once
Colour-only distinctions vanish, and some of your reviewers will read on a bad screen or on paper. This takes one minute and catches a figure that was unreadable to a third of your audience.
The conclusion is read early, not last
It is where a reviewer checks whether the paper claims more than it showed. A conclusion that restates the abstract's widest claim, after an evidence section that narrowed it, is where the two get compared.
The half of the list you cannot check yourself
Everything above is a check you can run alone. What is left is the reading: specialists go through the draft against the guidelines of the venue you are aiming at and return the objections with locations, while the file is still open.
Choosing between the two big venues is ICLR vs NeurIPS. When the reviews come back, the reply is how to write a rebuttal.
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.