Guides
Written for the questions authors actually ask
Where we can answer with measurement, we do, and the numbers come from 697 real ICLR and NeurIPS decisions. Where the honest answer is craft knowledge, the page says so.
The reviews are back
NeurIPS review scores explained
Five scored fields doing five different jobs. What each one asks, which of them a rebuttal can still move, and why confidence decides how much a bad score costs you.
ICLR review scores explained
The scale has no 2, 4, 7 or 9, and the gaps are deliberate. What each rating means, and why the whole conference happens in the one point between 5 and 6.
ICML review scores explained
Three editions, three different scales. A 4 was a borderline reject in 2024, an accept in 2025 and a weak accept in 2026, and the page ICML itself leaves at the top of the results explains the oldest of them.
How to write a rebuttal
Papers whose scores rose in the ICLR rebuttal were accepted at 55.7%, against 7.8% for those that did not. Which objections respond, whose review to answer first, and what to cut.
Why ML conference papers get rejected
Journal advice does not describe NeurIPS. Across 697 conference decisions there is one band, half a point wide, where the outcome stops tracking the manuscript at all.
Before you submit
Review my paper before submission
Three layers of checking and only one of them can be bought. What nobody but you can settle, what a script settles, and what needs someone to read the argument.
Where to submit your ML paper
Not a directory. Name the strongest thing your paper has, match it to the room that knows how to score it, and see where the bars actually sit.
The pre-submission checklist
Ordered by how far each failure kills you. Desk rejection first, because it costs the cycle and returns no feedback, then what reviewers punish, then the first five minutes.
ICLR vs NeurIPS
The papers NeurIPS turns down score where ICLR's accepted papers sit. Where the two bars really are, measured on 494 real decisions.
ICML vs NeurIPS
The review forms are almost the same document, down to a 1 to 6 overall score. What actually separates these two is the calendar: January against May.
CVPR vs ICCV vs ECCV
Only two of the three run in any given year, so the three-way choice is not one you ever make. Plus the CVPR rebuttal rule that breaks habits brought from NeurIPS.
ACL vs EMNLP
You do not choose between them at submission. Both take papers only through ACL Rolling Review, so you submit to a cycle, read your reviews, then commit to a venue.
ACL Rolling Review explained
The 18.9% everyone quotes counts papers that never reached a decision. Commit, and roughly seven in ten got in. Plus the three days you actually have to choose, and the score scale ARR mis-publishes on its own site.
AAAI vs IJCAI
One puts a labelled AI reviewer on every paper. The other charges 100 dollars a submission and bans LLMs from reviewing entirely. A lukewarm review kills at one venue and is survivable at the other.
How AAAI review actually works
Phase 1 arrives with the full reviews attached and no right of reply. What each stage lets you do, and why the whole defence is 2,500 characters against as many as five reviews.
COLM explained
Its own programme chairs say COLM papers count less because it is unranked, and that the quality is not lower. What that trade buys, and why the March deadline is a fork rather than a fallback.
AISTATS vs UAI
Not rivals, consecutive slots: AISTATS decides in January and UAI closes five weeks later. Both publish their full review forms, and a 4 means opposite things on the two of them.
WACV vs BMVC
WACV round one is not a free shot: a reject there cannot be resubmitted to round two. What it does buy is the only revision in the CVF calendar that the same reviewers then judge.
Conference or journal
The CRA says conferences are on par with or preferred to journals. Spain caps conference papers at three of five. Both are current, and which one applies to you depends on who reads your CV.
When the answer is no
After a NeurIPS rejection
Rerun the review and 82% of rejected papers are rejected again, which is the number nobody quotes. And this cycle ICLR closes six days before the decisions land, so the advice everyone repeats no longer works.
After an AAAI Phase 1 rejection
It is not a desk rejection: two humans and one AI read the paper, and all three reviews arrive with a notice you cannot answer. Only two of them carried a rating, and the first deadline left after the 24th is five days away.
Desk rejection explained
In 2026 the thing most likely to kill a paper before review is the author list, not the PDF. What each venue actually checks, why the abstract deadline is the real one, and the twenty minutes that catches the rest.
After a CVPR rejection
Nobody asks what you changed, which is the trap. ECCV counts self-plagiarism as plagiarism and bars an improved method from the supplementary, and the window to reach it was thirteen days.
TMLR explained
It has abolished not novel enough in writing, and made claims against evidence the one thing it enforces harder. Sort your own meta-review into what that saves and what it will not, beside the certification counts: four Outstanding awards in the journal's history.
On the tool itself
How accurate is AI peer review
91.68% against 697 decisions those committees had already made, 95.96% on ICLR 2025 with one false accept in 149 rejects, and the weakest set left in the table.
What AI peer review cannot do
Eight objections the field raises, put as sharply as we can put them and answered by the people building one of these. Several of them are simply correct.
AI review vs human review
NeurIPS sent the same papers through two committees twice, and about half of one's accepts were rejected by the other. What that does to every claim about AI reviewers, including ours.
Or skip the reading
The matcher reads what your paper is about and ranks the venues that publish it. Free, no account, and nothing you paste is kept.