Guide
How to write a rebuttal
A rebuttal has one job: move a number. Everything that does not move a number is decoration, and decoration costs you the word count you needed for the answer that would have worked.
Published 11 August 2026
Reviews are back and the paper is locked. The reply is the only thing you still control at this point. If you are reading this before you submit, the cheaper version of a rebuttal is a read of the draft while it can still change: have it read first
First, work out whether it is worth it
- Sorting 697 papers by how they read splits them into three groups. A rebuttal is worth a weekend in only one of them.
- If the reviews name a problem you can see in the paper, the reply is not what changes it. That is the group where 15 of 220 papers got in.
- If the reviews are close and contradictory, this is where a reply decides things. That is the band where 64 of 154 papers were accepted.
- Which one you are holding is usually obvious. It shows in whether the objections agree with each other.
What responds to what
Five objections, and how much each one will give you
Rebuttal windows are short and word limits are real. Spend them in the order below, not in the order the reviews arrived.
| The objection | How much it moves | What actually moves it | Worth the words |
|---|---|---|---|
| A factual misreading | Most | A pointer to where the paper answers it, plus an admission that it was not findable. This is the cheapest score you will ever recover. | Yes, always answer |
| A missing baseline or ablation | A lot, if you run it | The table. A promise to add it later is worth a fraction of the number itself, and reviewers have learned to discount promises. | Yes, if it fits the window |
| A disputed technical step | Some | The derivation, the counterexample or the experiment that settles it. Nothing rhetorical works here. | Only with the answer |
| Not novel enough | Least | Naming the thing that was not possible before, in one sentence. Usually a disagreement about values, and rarely settled inside a rebuttal. | Rarely |
| Hard to follow | Almost nothing now | A revision. At NeurIPS this is a rating of the submitted PDF, and that file is frozen. Answer it for the record and spend your words elsewhere. | No |
Craft knowledge, not measurement. We hold conference decisions, not reviewer threads, so nothing in this table is a success rate. The measured part is below.
What the record shows
Rebuttals do work, and mostly they do nothing
ICLR publishes every review and every reply, so the effect can be counted rather than guessed. Across ICLR 2024 and 2025, more than 19,000 papers and 74,000 reviews, three quarters of review scores did not move at all. The quarter that did decided a lot.
| What happened to the score | Papers | Share | Accepted |
|---|---|---|---|
| It went up | 5,807 | 50.8% | 55.7% |
| It did not move | 5,247 | 45.9% | 7.8% |
| It went down | 377 | 3.3% | 8.0% |
ICLR 2025, 11,431 papers, against an overall acceptance rate of 32.1%. ICLR 2024 gives the same shape: 57.6% accepted where scores rose, 12.4% where they held. Counted from the OpenReview record by Kargaran et al. These are ICLR figures and nothing here is ours.
Moving one score is worth roughly seven times the odds
Papers whose scores rose were accepted at 55.7%, against 7.8% for papers whose scores stayed where they were. That gap is the entire argument for writing a rebuttal at all. It is also not a promise: 44% of the papers whose scores went up were still rejected.
Almost nothing moves, so a move is a real signal
Three quarters of individual review scores were unchanged after the rebuttal, and score decreases were about one review in a hundred. The fear of talking a reviewer into lowering your score is real but rare. Silence is the more expensive mistake.
The movement is concentrated at the border
The most frequent score changes were 5 to 6, 6 to 8, and 3 to 5. Rebuttals move papers across the acceptance line, not from bad to good. If your scores are uniformly low, the record says this is the group rebuttals help least: for rejected and withdrawn submissions reviewer disagreement barely converged at all, under 7%, against 41 and 48% for orals.
A strong score is not a safe score
More than 40% of the papers sitting in the top 5% before the rebuttal were displaced from it afterwards, because other authors argued and they did not. If your reviews are good, the rebuttal is where you protect a position rather than where you relax.
Discussion beats a single long reply
Reviews whose scores rose show more back-and-forth turns and higher participation from both sides than reviews that stayed flat. And timing matters: replies posted very early or at the last minute were less often followed by a score increase than replies in the middle of the window, where nearly a third of opening messages were.
Triage
Answer the reviewer who can still be moved, not the one who annoyed you
Rank the reviews by how much a good answer would change, which is almost never the order in which they made you angry.
Low score, low confidence: start here
A reviewer who marked themselves unsure is telling you, in the conference's own wording, that they may have missed the central parts of the submission. That is a comprehension failure, and comprehension failures are the ones a paragraph can fix. This is the most winnable review in your batch and it is routinely answered last, because it is usually also the vaguest and therefore the most irritating.
Low score, high confidence, specific objection: answer it or accept it
Someone who names the exact step they do not believe has given you a gift, however it feels. Either you have the derivation or the experiment, in which case this is the highest-value answer on your list, or you do not, in which case no amount of writing helps and the words belong to another reviewer.
Low score, high confidence, no specifics: write for the chair
You cannot answer an objection that was never made. What you can do is make the absence visible: answer what little was raised, completely and without complaint, and let a review with no substance sit next to a reply with plenty. An area chair reading the thread draws the conclusion themselves, which is worth more than you drawing it for them.
Your supporter needs two lines, not none
The reviewer who liked the paper still asked questions, and leaving them unanswered while you fight with someone else reads exactly as it looks. They are also the person who will defend you in the discussion, and they can only do that with material you gave them.
The shape
What a reply that works looks like
Reviewers read these under time pressure, often on a phone, often having forgotten the paper. The structure below is built for that reader rather than for a careful one.
- Open with what changed, in three lines. New results first, corrections second. If you ran the experiment they asked for, the number goes in the first sentence, not in paragraph six. Many reviewers read only this far before deciding whether to re-open the paper.
- One block per reviewer, headed by their own words. Quote the objection in the reviewer's phrasing before answering it. It lets them find their point in a wall of text, and it proves you understood the criticism before disagreeing with it.
- Answer, then evidence, then the change. State the answer in one sentence. Give the number or the derivation. Say what you changed in the paper and where. In that order, every time: a reader who stops after the first sentence has still received the answer.
- Concede fast and completely where you are wrong. A clean concession costs a sentence and buys credibility for the disagreements that matter. Defending everything makes your strong answers look like more of the same.
- Put anything raised twice in a shared block at the top. Two reviewers with the same concern is the signal that the paper caused it. Answer it once, properly, where the chair will see it, rather than twice at half strength.
- Close by naming what you did not do, and why. An experiment you could not run inside the window is not a weakness if you say so plainly with the reason. It becomes one when a reviewer notices the silence.
The ones that cost papers
Six sentences to cut before you post
The reviewer clearly did not read the paper
Or any of its polite variants, including the one that opens with respectfully.
Even where it is true, it converts a fixable misunderstanding into a dispute with a person, in front of the chair, in writing, on a record that at ICLR stays public. If a reviewer missed something that is in the paper, the paper hid it. Say where it is, say you have made it easier to find, and move on in two lines.
We will add this in the camera-ready version
Used for the missing experiment rather than for a typo.
For small things it is fine and expected. For the central missing result it asks a reviewer to accept a paper on trust, after they have just told you they do not have enough evidence, which is the one currency you have already spent. If you can run it, run it. If you cannot, argue why the paper stands without it.
This is out of scope for this paper
Deployed against the concern the reviewer built their score on.
Scope is a real defence and it is the most over-used one. It works when the request would be a different paper and you can say what that paper would be. It fails when it is used to decline the question the reviewer cares most about, because from the other side it reads as a refusal rather than an argument.
As we already stated in Section 3
On its own, with nothing after it.
If a reviewer missed it, pointing at it again does not help them and the tone does the opposite. The version that works is the same pointer plus one sentence restating the content, plus what you have changed so the next reader finds it. Same information, and it reads as help rather than as a scolding.
This is standard practice in our community
Offered instead of a citation.
Often correct, and useless as written: your reviewer either does not know the convention or does not accept it, and an appeal to a consensus they are not part of settles nothing. Name two papers that do the same thing and the objection is over.
We respectfully disagree
Followed by a restatement of what the paper already said.
Disagreement is fine and sometimes necessary. Disagreement with no new information is a request that the reviewer change their mind because you would like them to. If you have nothing new to add on a point, either concede it or leave it, and use the space on a point where you do.
You cannot edit the paper. You can still watch the problem.
The reply goes in, the decision takes weeks, and work on your question keeps appearing the whole time. Name the topic and Gap Alerts email you when a paper covering the same ground lands, in time for the camera ready or for the next submission, whichever this becomes.
What the numbers on your review mean, field by field: NeurIPS and ICLR.
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.