Guide
ICML review scores: a 4 means three different things
ICML changed its overall scale in three consecutive editions, so the same digit reads Borderline reject in 2024, Accept in 2025 and Weak accept in 2026. Nobody has published that table; the page outranking this one is ICML's 2024 instructions.
Published 14 August 2026
Decoding a score you were just given? Then the file is with the area chairs and nothing on this site changes it now. If the draft is still open in front of you, the page that helps is the pre-submission checklist
The 497 is ICML's own count, from its post on reviewer LLM policy violations.
Find the edition before you read the number
- The 6 that tops the 2026 scale was only Weak Accept in 2024. The same digit, two different verdicts.
- Work out which edition produced your review, then read that year's instructions. Advice that names no year describes a form that no longer exists.
Three scales, three years
The scale changed under the same conference name
Not a revision of anchors: three different ranges, with the sub-scores and the confidence field appearing, disappearing and returning.
| Edition | Overall | The rungs, top to bottom | Sub-scores | Confidence |
|---|---|---|---|---|
| ICML 2024 | 1 to 10 | 10 Award quality, 9 Very Strong Accept, 8 Strong Accept, 7 Accept, 6 Weak Accept, 5 Borderline accept, 4 Borderline reject, 3 Reject, 2 Strong Reject, 1 Very Strong Reject. | Three, each 1 to 4: soundness, presentation, contribution. | Present, 1 to 5. |
| ICML 2025 | 1 to 5 | 5 Strong accept, 4 Accept, 3 Weak accept, 2 Weak reject, 1 Reject. | None on the main track. | None. Deleted as "difficult to interpret". |
| ICML 2026 | 1 to 6 | 6 Strong Accept, 5 Accept, 4 Weak accept, 3 Weak reject, 2 Reject, 1 Strong Reject. | Four, each 1 to 4: soundness, presentation, significance, originality. | Restored, 1 to 5. |
Each row from that edition's own page, read August 2026: ICML 2024, ICML 2025, ICML 2026, with the deletion and its reason in the 2025 area chair instructions. ICML 2026 is the newest edition we can confirm: the 2027 pages 404.
The folklore comes from one stale page ICML never took down
The 2024 instructions still sit on icml.cc under their own URL, which makes them the most authoritative-looking answer to a question they now answer wrongly. That is why search summaries say ICML uses a ten-point scale.
ICML 2026, the current form
What each of the six rungs says, in ICML's words
Quotes, not paraphrases: the anchors are what your reviewer selected against.
| Score | Label | ICML 2026 wording, verbatim |
|---|---|---|
| 6 | Strong Accept | "Technically flawless paper with exceptional impact on one or more areas of AI, with strong evaluation, reproducibility, and resources, and no unaddressed ethical considerations." |
| 5 | Accept | "Technically solid paper, with high impact on at least one sub-area of AI or moderate-to-high impact on more than one area of AI, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations." |
| 4 | Weak accept | "Technically solid paper that advances at least one sub-area of AI, with a contribution that others are likely to build on, but with some weaknesses that limit its impact (e.g., limited evaluation). Please use sparingly." |
| 3 | Weak reject | "A paper with clear merits, but also some weaknesses, which overall outweigh the merits. Papers in this category require revisions before they can be meaningfully built upon by others. Please use sparingly." |
| 2 | Reject | "For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility, incompletely addressed ethical considerations, or writing so poor that it is not possible to understand its key claims." |
| 1 | Strong Reject | "For instance, a paper with well-known results, unaddressed ethical considerations, or a poorly written paper where it is impossible to tell what the nature of its contribution is." |
Quoted from the main track form in the ICML 2026 reviewer instructions, read August 2026. The tinted rows are the two ICML asks reviewers to use sparingly.
Use sparingly sits on the hedges, not on the extremes
ICML attaches that instruction to 4 and 3 and to neither end of the scale: it discourages the safe middle, not the strong opinion. So a 4 or a 3 is not a shrug. It is the answer a reviewer picked after being told to avoid it, and the weaknesses they wrote down are load-bearing.
The accept rungs are checkable, the reject rungs are examples
A 6 needs technically flawless work and, in the same sentence, strong evaluation, reproducibility and resources with no unaddressed ethical considerations, and the 5 repeats that chain, so a thin evaluation caps you below both however good the idea is. Both reject anchors open with "For instance", so a 2 or a 1 need not match a listed case, and a rebuttal arguing that none of them fits is aimed at the wrong sentence.
The rest of the 2026 form
Four sub-scores out of 4, and the rules attached to them
The recommendation is one field. Two of the others oblige your reviewer to write something down.
| Field | Range or format | What ICML attaches to it |
|---|---|---|
| Soundness | 1 to 4 | 4 excellent, 3 good, 2 fair, 1 poor |
| Presentation | 1 to 4 | The same four anchors |
| Significance | 1 to 4 | The same four anchors |
| Originality | 1 to 4 | The same four anchors |
| Justifying a fair or a poor | Prose, required | "If you select "fair" or "poor" (indicating that the paper falls short of the standard), ensure that "Strengths and Weaknesses" include a clear justification of your rating." |
| How the four combine | Nothing published | No weighting, no formula, no stated relationship to the recommendation. Any page giving you one invented it. |
| Final Justification | After the rebuttal, new in 2026 | Says whether the rebuttal "changed your evaluation", and "is shared with the authors, AC, SAC, and PCs". |
Field wording from the ICML 2026 reviewer instructions, read August 2026. The tinted rows are the two that put a sentence in front of you, and a rating filed without that sentence is an incomplete review by ICML's own rule.
Same digits, different words
A 4 is not even one thing inside ICML 2026
The position track runs its own form on the same page, over the same range with relabelled rungs. Put NeurIPS 2025 beside them and the main track is the odd one.
| Score | ICML 2026 main track | ICML 2026 position track | NeurIPS 2025 |
|---|---|---|---|
| 6 | Strong Accept | Strong Accept | Strong Accept |
| 5 | Accept | Accept | Accept |
| 4 | Weak accept | Borderline accept. Please use sparingly. | Borderline accept. Please use sparingly. |
| 3 | Weak reject | Borderline reject. Please use sparingly. | Borderline reject. Please use sparingly. |
| 2 | Reject | Reject | Reject |
| 1 | Strong Reject | Strong Reject | Strong Reject |
Both ICML columns from the ICML 2026 reviewer instructions, NeurIPS labels from the NeurIPS 2025 reviewer guidelines, dated 2025 on purpose: the 2026 edition publishes no numeric anchors. Borderline describes how close the call is; weak describes the paper.
If you submitted a position paper, the table above is the wrong form
Its sub-scores are five other things, and its alternative views section is, verbatim, "mandatory. It must be in the main body of the paper (not an appendix) that describes and addresses one or more credible (not strawman) positions that are opposed to the paper's position."
The two forms compared in full, and the calendar that decides
The number nobody has
There is no score-to-outcome table for the 1 to 6 scale
Not withheld: it does not exist. No published distribution maps ICML scores to decisions on either recent scale. The only real numbers sit on a scale ICML retired.
| ICML 2023 outcome | Mean average post-rebuttal score, retired 1 to 10 scale |
|---|---|
| Rejected | 4.32 |
| Accepted as poster | 5.93 |
| Accepted as oral | 6.82 |
| Outstanding Paper Award | 7.72 |
From Su et al., The ICML 2023 Ranking Experiment, across 6,538 submissions, on the scale ICML retired after 2024. These numbers cannot be read against a 1 to 6 review. Our own scoring has never met an ICML decision either, so we supply no replacement.
This is why the score predictors ranking for your query are broken
They were fitted on the only public data that exists, the 1 to 10 era, and they still take a 1 to 10 input. Feed a 2026 review in and a 4 arrives as a borderline reject when it is a weak accept.
ICML tells its own chairs not to read the average, and tells you too
Area chair instructions: recommendations and meta-reviews "should be informed by the review content, as opposed to just relying on average (or other aggregate) numerical scores." The peer review FAQ, to authors: "the average score should not be taken as a direct indication of the final decision on the paper." In place of a cut-off there is a sentence: papers "technically sound, well-written, non-redundant with previous research, and useful to at least some fraction of the ICML community should be accepted".
Quotes and criterion from the ICML 2026 area chair instructions, which also bar capacity arguments, and the peer review FAQ. The 26.6% is ICML's own, from two postings that disagree on the totals: 23,918 submissions and 6,352 accepted on 30 April 2026, 24,661 and 6,552 on 7 July, confirmed by the communications chairs' retrospective.
The score is filed. The topic is not.
In January 2026 ICML gave authors AI feedback on roughly 4,500 papers, outside peer review, then published the survey: of 869 respondents, 92.1% would use it again. That job needs a PDF you can still change. While this one sits with the area chairs, name the topic it is about and Gap Alerts email you when a paper covering the same ground appears.
A review costs credits and returns objections with locations, not a prediction: what AI peer review cannot do. When the reviews arrive: how to write a rebuttal. Before the next one goes out: the pre-submission checklist. The other big form: NeurIPS review scores explained. Programme figures from the ICML retrospective on its AI paper assistant programme, read August 2026. ICML's own programme on its own submissions, not this service.
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.