How to write an ICLR 2027 rebuttal
Acadwrite team · Last updated 2026-11-06 · The official iclr.cc pages always take precedence
ICLR 2027 moved the overall rating to four levels, 1–4. Authors reply under each review on OpenReview, reviewers can follow up and change scores, and the paper can be revised during the discussion.
For general conference rebuttal advice, see How to write a conference paper rebuttal.
1. What you can do during the discussion
| Question | Answer |
|---|---|
| Can I revise the paper? | Yes. Until Nov 18 you can upload revised PDFs as often as needed, and a pdfdiff against the original is generated. ACs and reviewers may ignore changes that differ too much from the original. |
| How many pages? | The main text goes from 9 to 10 pages; the extra page is for new results. Papers over the limit are desk-rejected. References and acknowledgements don't count. The appendix is unlimited, but reviewers needn't read it. |
| Figures in a comment? | No. OpenReview comments don't render images. Put new figures in the revised PDF and give their location in the reply. |
| Tables and equations in a comment? | Yes. Comments render Markdown tables and LaTeX. |
| How long can a reply be? | Each comment has a length limit; in 2025 it was 5,000 characters. The number of comments is unlimited, so split long replies. |
| Can I change the title or abstract? | The Author Guidelines conflict: the FAQ says yes, the author-order section says the title can't change before the discussion ends. Keep the title; mention any abstract change in your reply. |
| Can I share a code link? | Yes. Post an anonymous repository link in a comment visible only to reviewers and ACs. Demo and video links must be fully anonymous, on a host that doesn't track visitors. |
| Can I contact the AC privately? | Yes. Set the comment's readers to Area Chairs only. Issues the AC can't resolve go to the Program Chairs by email, subject prefixed [Authors-to-PCs]. |
2. The rating scale: four levels in 2027
ICLR 2027 uses a new review form. The overall rating has four levels:
| Score | Form wording | In short |
|---|---|---|
| 4 | Clear acceptance | Clear accept |
| 3 | Weak acceptance | Leaning accept |
| 2 | Weak rejection | Leaning reject |
| 1 | Clear rejection | Clear reject |
Sub-scores and confidence
| Item | 1 | 2 | 3 |
|---|---|---|---|
| Soundness | Significant concerns that do not appear easy to address | Minor or fixable concerns | No substantive concerns |
| Presentation | Difficult to follow despite careful reading | Generally understandable, but requires substantial effort in places | Clear and easy to follow |
| Informativeness | Learned little or nothing new | Learned something interesting or useful | Gained a substantially deeper understanding of the problem |
- 2026 used 0/2/4/6/8/10 and 2025 used 1/3/5/6/8/10. The new scale has no "marginally above/below" level.
- Contribution was replaced by informativeness, and all three sub-scores went from four levels to three.
- Confidence went from five levels to four; 4 means substantial direct research experience on the topic.
- The review text is split into critical strengths, critical weaknesses and questions for authors.
3. This year's initial score distribution
We tallied every public initial review on OpenReview: 40,462 papers and 125,059 reviews, with about nine in ten papers having three reviewers. Data as of Nov 6; a few reviews may already have changed during the discussion.
Where your average ranks within your area
Each cell is the share of papers in the area whose average is at or above that score; areas are sorted by mean score. With three reviewers, 2.00 is 2/2/2 or 1/2/3, 2.33 is 2/2/3, 2.67 is 2/3/3 and 3.00 is 3/3/3.
| Area | Mean | ≥2 | ≥2.33 | ≥2.5 | ≥2.67 | ≥3 |
|---|---|---|---|---|---|---|
| All areas | 2.06 | 70% | 38% | 17% | 15% | 4.1% |
| Generative Models | 2.15 | 77% | 47% | 22% | 20% | 5.4% |
| Learning Theory | 2.12 | 69% | 44% | 25% | 23% | 8.9% |
| Applications: Robotics | 2.12 | 76% | 43% | 19% | 17% | 4.4% |
| Applications: CV / Audio / Lang | 2.11 | 75% | 42% | 18% | 16% | 3.8% |
| Optimization | 2.08 | 70% | 41% | 20% | 18% | 5.5% |
| Infrastructure / Systems | 2.08 | 70% | 39% | 19% | 18% | 5.6% |
| Foundation Models / LLMs | 2.07 | 71% | 38% | 17% | 15% | 4.3% |
| Datasets & Benchmarks | 2.06 | 69% | 38% | 17% | 15% | 4.2% |
| Applications: Neuro & CogSci | 2.05 | 70% | 39% | 14% | 14% | 3.7% |
| Reinforcement Learning | 2.04 | 68% | 37% | 16% | 15% | 3.9% |
| Probabilistic Methods | 2.04 | 67% | 37% | 16% | 15% | 4.1% |
| Representation Learning | 2.04 | 66% | 35% | 16% | 15% | 4.7% |
| Alignment / Safety / Privacy | 2.03 | 69% | 35% | 14% | 13% | 3.0% |
| Graphs & Geometry | 2.02 | 67% | 34% | 15% | 13% | 3.9% |
| Other Topics in ML | 2.01 | 67% | 33% | 13% | 12% | 3.1% |
| Applications: Physical Sciences | 2.00 | 64% | 33% | 14% | 12% | 3.4% |
| Transfer / Meta / Lifelong | 1.99 | 65% | 31% | 12% | 11% | 2.3% |
| Time Series & Dynamics | 1.99 | 63% | 31% | 12% | 10% | 2.5% |
| Causal Reasoning | 1.97 | 60% | 33% | 12% | 12% | 2.8% |
| Neurosymbolic & Hybrid AI | 1.93 | 59% | 29% | 11% | 10% | 2.8% |
| Interpretability / XAI | 1.92 | 56% | 28% | 12% | 10% | 2.8% |
An ICLR 2027 SAC publicly shared figures for the 200+ papers in their batch that match: mean 2.06, median 2.00, quartiles 1.67 and 2.33.
With three reviewers, one moving from 2 to 3 lifts the average by 0.33: 2/2/3 is in the top 38%, 2/3/3 in the top 15%.
4. Do scores change after the rebuttal?
A paper analysing public ICLR 2024–2025 review data (third-party research, old scale) measured changes before and after the discussion:
For reference: 2025 accepted 32% of 11,603 submissions; 2026 about 27.4%.
5. Writing the replies
General comment first · Post one comment visible to all reviewers listing the revision's changes and new experiments. Answer concerns shared by several reviewers there.
Then reply point by point · Restate each concern in one line, then give the answer and the evidence (numbers, tables, location in the revision).
Key concerns first · Reviewers are asked to state the one or two key reasons behind their recommendation. Answer those first and group minor points.
Reply early · Reviewers update their reviews after the discussion. Replies posted in the last day or two rarely get a follow-up or a score change.
Everything is public · After decisions, papers are de-anonymised, and rejected and withdrawn papers stay on OpenReview with author names.
Finish your own reviews · Authors on 3 or more submissions must review at least 6 papers. Those without complete, adequate reviews by the rebuttal stage risk desk rejection.
6. Figures, tables and equations
An example reply (numbers are placeholders):
- Number new figures and tables Figure R1, Table R1 to keep them apart from the original.
- Colour-coding is not required; the guidelines only ask that changes be communicated to reviewers and the AC. The usual practice is to colour new text and list locations in the general comment.
7. AI-use rules
- Papers must include an AI use statement, which doesn't count toward the page limit. Update it if new results relied substantially on an LLM.
- Serious fabrication, plagiarism or misrepresentation caused by an LLM violates the Code of Ethics and can lead to desk rejection.
- Hiding prompts in a paper to get favourable AI reviews counts as collusion (2026 policy).
There is no separate rule on disclosing AI use in rebuttal text.
8. Writing the rebuttal with co-authors
Rebuttals are usually split among co-authors and checked by the corresponding author. In an Acadwrite workspace:
The PDF itself is still revised in your LaTeX project.
FAQ
What if a reviewer never responds?
A few days before the deadline, post a short follow-up under that review listing the changes made for them. If there's still no reply, explain in a comment visible only to the AC.
Is it worth replying with 1/2/2?
1/2/2 averages 1.67, and about 71% of papers average higher, so reaching acceptance through the rebuttal is unlikely. A reply can still correct factual misreadings, or the time can go into resubmitting elsewhere.
What happens if I withdraw?
You can withdraw any time before decisions. Withdrawn papers stay in the public withdrawn section, de-anonymised immediately, with reviews still visible.
Sources
- OpenReview: ICLR 2027 Conference (public reviews)
- ICLR 2027 Dates
- ICLR 2027 Author Guidelines
- ICLR 2027 Reviewer Guidelines
- ICLR 2027 Call for Papers
- ICLR 2027 AI Policy for Authors
- OpenReview: formatting comments with Markdown and LaTeX
- ICLR 2025 Fact Sheet
- ICLR Blog: ICLR 2026 retrospective
- ICLR Blog: 2026 discussion-period security incident
- ICLR Blog: LLM and prompt-injection policy
- Kargaran et al., analysis of ICLR rebuttal score changes (third-party)