How to write a rebuttal for an ML conference

The area chair writes the recommendation from the reviews, the rebuttal and the discussion. A good rebuttal does three things: it fixes misreadings, it supplies missing evidence, and it shows the area chair you took the criticism seriously. It rarely changes a reviewer's view of how significant the work is.

So spend your time where scores can move: factual errors and missing experiments raised by reviewers sitting near the bar.

1. Read every review once, then put it away for an hour. Answer nothing yet.
2. Copy every distinct concern into one table: one row each, with the reviewer, a short quote and their score and confidence.
3. Merge duplicates. A concern raised by two reviewers is the main issue, so answer it in a general response.
4. Tag each row by type (table below) and by cost: minutes, hours, or not possible in the window.
5. Find the persuadable reviewers: borderline scores with real questions. They get the most effort.
6. Start any new experiments today. Text can be written while they run.

Use the same four beats every time, so reviewers and the area chair can skim:

1. Restate the concern in one line, so it's clear what you're answering.
2. Answer directly in the first sentence: “Yes, we ran this,” or “This is covered in Section 4.2.”
3. Give evidence: a number, a table, a quote with section and line, or a citation.
4. Say what changed in the paper, or will change in the camera-ready.

The harsh, low-score, high-confidence review is written for the reviewer, but your answer is written for the area chair. Stay factual and short, and don't argue about tone. If the review contains a clear technical error, show it with evidence once and move on. If it ignores something in the paper, quote the passage.

If a review is clearly careless (for example, it doesn't engage with the paper at all), many venues have a confidential channel to the area chair. Use it sparingly and give specifics.

The mechanics differ a lot. Check the current cycle's author guide, since these change every year.

Adapt the structure to your venue's format and length limit.

We thank the reviewers for their time. We are glad they found [strength they agreed on].

Summary of changes (new results in the revised PDF are marked in blue):
1. [New experiment / ablation]: [one-line result], Section X, Table Y.
2. [Clarification]: [what changed], Section Z.

--- Reviewer [ID] ---
Q1 (restated briefly): "[their concern in one line]"
A1: [Direct answer first.] [Evidence: number, table, quote with section/line.] [What we changed.]

Q2: ...
A2: ...

We hope this addresses the concerns about [main issue]. We are happy to clarify anything else.

• Promising experiments "for the camera-ready" instead of running them now, when the window allows it.
• Answering the easy questions at length and the main concern in one vague line.
• Arguing with a reviewer's taste ("we respectfully disagree that this is incremental") without new evidence.
• Breaking anonymity in a double-blind venue, e.g. a link to a non-anonymous repository.
• Ignoring the length limit or formatting rules. Some venues tell area chairs to disregard non-compliant rebuttals.
• Writing to the reviewer who will never move, instead of the two who might.