Writing a NeurIPS/ICML Rebuttal

A NeurIPS- or ICML-style rebuttal is a distinct genre from a journal response-to-reviewers letter, and the differences are not stylistic — they are hard mechanical constraints set by the submission platform and the conference’s double-blind policy. You get a fixed, short window (commonly about a week, though exact dates vary by conference and year), a per-review character limit enforced by the software rather than editorial discretion, no ability to revise your submitted PDF, and an absolute anonymity rule that a journal cover letter never has to worry about. This guide covers those mechanics specifically. For the underlying skill of deciding what to respond to when you don’t have space for everything, see Prioritizing Major vs. Minor Reviewer Comments. For how CS conference authorship itself differs from journal authorship, see Authorship Credit in Computer Science: Conference Papers vs. Journal Papers. Neither of those pages covers OpenReview’s specific submission mechanics, which is the gap this one fills.

What ‘rebuttal’ means in this context

In CS conference peer review (NeurIPS, ICML, ICLR, and most venues that follow the same model), the ‘rebuttal’ or ‘author response’ is a single round of written replies authors post to each reviewer’s comments during a fixed discussion window, after initial reviews are released and before final decisions. It is not a revised manuscript and not a negotiation with an editor — it is a structured, time-boxed exchange conducted inside the submission platform (almost always OpenReview for NeurIPS, ICML, and ICLR) that area chairs and reviewers read before finalizing scores and recommendations.

This is a materially different format from a journal response-to-reviewers letter, which is typically a standalone document with no hard length limit, submitted alongside a revised manuscript, addressed to an editor who mediates between authors and (usually anonymous-to-authors) reviewers over one or more rounds. See Single-blind vs. double-blind vs. open peer review for how the anonymity model itself differs across venues, and How to Write a Response to Reviewers Letter for the journal-side equivalent of this guide.

The double-blind anonymity rule is absolute, and it applies to you too

Double-blind review means neither reviewers nor authors know each other’s identities during review. That constraint doesn’t relax during the rebuttal period — if anything, it tightens, because a rebuttal is free-text prose you’re writing in real time under deadline pressure, which is exactly when an accidental deanonymizing slip is most likely.

No identifying links. NeurIPS’s own author FAQ is explicit: do not use links in any part of the rebuttal. The one narrow exception is code a reviewer specifically requested, which may be shared as an anonymized link, and even then it typically has to go to the area chair via an official comment rather than posted openly to the reviewer thread — check your specific conference’s current guidance before relying on this exception.
No self-identifying phrasing. Avoid “in our prior work [X]” where X is your own unblinded citation, references to your institution, lab, grant numbers, or anything else that narrows down who you are. This is the same discipline your original anonymized submission required — it doesn’t lapse just because you’re now writing informally under time pressure.
Anonymized supplementary evidence only. If you want to show a reviewer a new plot or table, it typically has to go through the platform’s own comment/attachment mechanism if one exists, not an external anonymous file host, unless the conference’s guidance explicitly allows it.

A rebuttal that breaks anonymity can be treated as a policy violation independent of the paper’s scientific merits, so this isn’t a stylistic nicety — treat it as a hard constraint on every sentence you draft.

The window is fixed and short — plan around days, not weeks

Unlike a journal revision deadline, which is often weeks or months and negotiable with the editor, the CS conference rebuttal period is a fixed calendar window set centrally for every paper at the venue, commonly on the order of about a week (exact dates are published on each conference’s own “Important Dates” page and vary year to year — always check the current cycle’s page rather than assuming last year’s schedule). There is generally no individual extension: the platform closes the response field to everyone at once.

Practical consequences of a fixed, non-negotiable window:

You typically don’t know your reviews until the rebuttal period opens, so there’s no way to start drafting before then — budget the entire task, from first read of the reviews to final submission, inside that single week.
If you have multiple co-authors, agree on a division of labor (who drafts the response to which reviewer, who does a final anonymity check) before the window opens, not after reviews land, because there’s no slack to recover lost days.
Because most platforms let you save drafts and keep editing until the deadline, submit an adequate response early and keep refining it, rather than waiting until the final hours to post a first draft that then has no time for a second look.

OpenReview mechanics: what you’re actually filling in

NeurIPS, ICML, and ICLR all run peer review through OpenReview, and the rebuttal is submitted as one or more “official comment” responses tied to the corresponding review, not as an uploaded document. A few mechanics that differ from a journal submission portal:

Per-review responses, not one combined letter. You generally write a separate response addressed to each reviewer rather than a single combined document, though platforms typically let every reviewer on the paper see every response, so you can point one reviewer to a point made in a response to a different reviewer instead of repeating yourself.
Character limits, enforced by the form itself. Limits are set per conference and can change year to year; recent published NeurIPS guidance sets the per-review response cap at 10,000 characters, and recent published ICML guidance sets it at 5,000 characters per response. These are hard limits enforced by the submission field, not soft editorial guidance — the form will not accept text past the cap, so verify the current year’s exact limit on the conference’s own author-instructions or FAQ page before you start drafting, and draft with a running character count rather than discovering the limit only when you try to submit.
Plain text with limited markdown, no file uploads. Recent NeurIPS guidance describes the response field as plain text with the subset of markdown formatting OpenReview supports (for example, basic emphasis and lists) — you generally cannot attach a revised figure, a new PDF, or supplementary files directly to a rebuttal comment the way you could attach a cover letter and revised manuscript to a journal portal.
Delayed visibility. Some conferences reveal all rebuttals to reviewers and area chairs only once the rebuttal window closes, rather than showing each response the moment you post it — check your venue’s current policy, since this affects whether you should expect any back-and-forth before decisions are finalized.

You cannot revise the paper itself

This is the single biggest structural difference from a journal revision, and it trips up authors who are used to the journal cycle: at most double-blind CS conferences, you cannot upload a revised PDF or updated supplementary material during the rebuttal period. Recent NeurIPS author guidance is direct about this — no revisions are allowed to the paper or the supplementary material until the camera-ready stage, which only happens after acceptance. Whatever is in your original submitted PDF is what reviewers and area chairs evaluate; the rebuttal text is the only thing you get to add.

New experimental results are typically allowed inside the rebuttal text itself (as numbers, small tables, or a described method), but recent NeurIPS guidance frames this narrowly: the original submission remains the basis for the acceptance recommendation, and the rebuttal exists to clarify reviewer and area-chair questions, not to substantively extend the paper’s contribution. Practically, that means new numbers in a rebuttal are useful for answering “did you check X” or “how sensitive is this to Y,” but they are not a substitute for having run the ablation before submission, and reviewers are entitled to discount a result they can’t independently verify because it isn’t in the reviewed PDF.

ICML’s published guidance narrows this further on the discussion side: your rebuttal response should contain your complete argument up front, because any later discussion during the remaining window is meant to address points already raised, not to introduce new arguments after the fact. Don’t hold points in reserve expecting a second round — there often isn’t one, and even where limited back-and-forth is possible, it is not the place to raise something you left out of your main response.

What actually moves a score, and what doesn’t

Rebuttals are read by people who have already formed an initial judgment and are deciding whether to revise it, inside a discussion period that is itself time-limited for the reviewer, not just for you. A few patterns that hold across NeurIPS/ICML-style venues:

Correct factual errors first and plainly. If a reviewer misread a result, misunderstood the method, or cites a number that isn’t what your paper says, state the correction directly, point to the exact section/table, and move on. This is the highest-value category of rebuttal content and the one most likely to actually change a score.
Answer the specific question asked. A reviewer question buried in a long paragraph of “weaknesses” still needs a direct answer; don’t let it get lost inside a general response to the review as a whole.
Don’t relitigate a subjective judgment as if it were a factual dispute. “Reviewer thinks the contribution is incremental” is a judgment call that a clarification alone may not resolve; over-arguing it can read worse than a concise, evidence-based case for significance.
A reviewer who doesn’t respond to your rebuttal hasn’t necessarily ignored it. Discussion periods are short for reviewers managing several papers at once; area chairs typically read all rebuttals when forming a recommendation even where individual reviewers stay silent. A missing reply is not evidence your response failed.

For the general skill of triaging which comments deserve full treatment when you’re short on space (a problem shared with journal response letters but especially acute under a hard character cap), see Prioritizing Major vs. Minor Reviewer Comments When Time or Word Count Is Limited.

A practical checklist

Confirm this year’s exact rebuttal window dates and per-response character limit on the conference’s own author-instructions/FAQ page — do not assume last year’s numbers carry over.
Re-read your own submission before drafting; you’re arguing about a fixed PDF, and citing it precisely (section, equation, table, line number if your venue’s format allows it) strengthens every reply.
Draft against a running character count, not after the fact — a response you have to cut at the deadline is worse than one paced correctly from the start.
Do one final pass for anonymity before submitting: no links except an explicitly permitted anonymized-code exception, no self-citation phrasing that unblinds you, no institutional identifiers.
Address every reviewer’s substantive points, even briefly — a reviewer whose concern goes unmentioned has no way to know you considered it.
Don’t hold back an argument for a hoped-for second round; the platform may not reveal responses immediately, and later discussion, where it exists at all, is generally meant to resolve points already on the table, not introduce new ones.
Coordinate with co-authors on who owns which reviewer’s response before the window opens, since there’s no slack to renegotiate roles once the clock starts.

Frequently asked questions

How long is the NeurIPS or ICML rebuttal period?

It’s a fixed window set centrally for the whole conference, commonly on the order of about a week, but the exact dates are published on that year’s conference website and change annually — always check the current cycle’s official dates page rather than assuming a prior year’s schedule applies.

Can I upload a revised PDF during the rebuttal period?

Generally no. Recent NeurIPS guidance is explicit that no revisions to the paper or supplementary material are allowed until the camera-ready stage, which happens only after acceptance. Check your specific conference’s current policy, since this is exactly the kind of detail worth confirming rather than assuming.

Is there a character or word limit on the rebuttal?

Yes, and it’s enforced by the submission form itself, not by editorial discretion. Recent published limits have been 10,000 characters per review response at NeurIPS and 5,000 characters per response at ICML, but these figures are set per conference per year — verify the current limit on the official author-instructions page before drafting.

Can I include new experiments in a rebuttal?

Often yes as described text/numbers, but with an important caveat: recent NeurIPS guidance states the original submission remains the basis for the acceptance recommendation, and new results in a rebuttal are meant to clarify reviewer questions rather than substantively extend the contribution. You cannot attach new figures or a revised PDF to demonstrate them — only what fits in the text response.

What happens if I include an identifying link or detail in my rebuttal?

It risks violating the venue’s double-blind policy independent of your paper’s merits. NeurIPS guidance is direct that links generally should not appear in rebuttal text at all, with a narrow, conference-specific exception for reviewer-requested code shared as an anonymized link through an official channel. When in doubt, leave it out and describe the result in words instead.

Do reviewers have to respond to my rebuttal?

No. Reviewers are not generally obligated to reply during the discussion period, and area chairs typically read all rebuttals when finalizing a recommendation even where a specific reviewer stays silent. Silence from one reviewer is not a reliable signal about whether your response was persuasive.

Related CASRAI guides

This guide describes common practice at NeurIPS-, ICML-, and ICLR-style double-blind conferences as documented in each venue’s own current author FAQ/instructions pages. Character limits, exact dates, and specific platform behavior are set independently by each conference for each year — always confirm current details on your target conference’s official author-instructions page before relying on any specific figure here.