Why ML conference papers get rejected

The abstract says the method improves generalisation. The experiments show it improves accuracy on three datasets in one domain at one model size.

This is the cheapest rejection to earn and the cheapest to avoid, because the fix costs you nothing that matters. Nobody is asking for more experiments. They are asking the sentence to describe the experiment you ran. A claim narrowed to what you actually demonstrated is not a weaker paper, it is a paper that survives its own evidence section, and the reviewer who was going to write this sentence now has nothing to write.

Read your own abstract as an adversary once, with a pen, and mark every noun that promises a scope. Then check each one against a figure. The ones with no figure behind them are the objections you are choosing to ship.

No comparison against the method everyone in this subfield would reach for first, or a comparison against a version of it that was not tuned.

A reviewer cannot tell an unlucky omission from a convenient one, and is not required to try. The damage is not the missing row in the table, it is what the absence implies about the rows that are there: if this one was left out, the reader now has to wonder about the others, and the whole evidence section loses the benefit of the doubt.

If a baseline genuinely does not apply, say so in the paper in one sentence with the reason. An explained absence costs nothing. An unexplained one costs the section.

The related work is thorough, the method is described in full, and after both the reviewer still cannot say in one sentence what this paper did that the previous one did not.

This is almost never an absence of novelty. It is novelty that was never stated, because by the time you write the paper the delta is so obvious to you that it reads as not worth a sentence. It is worth a sentence. It is worth the sentence that opens your contributions list, and it should name the thing that was not possible before.

The tell is a related work section that describes neighbouring papers accurately and never says what is wrong with them, or what they could not do. A survey paragraph is not a positioning paragraph.

The system has four new parts and one number at the end, so there is no way to know which part earned it, or whether one of them is doing nothing.

Missing ablations are read as a hidden negative result, and often correctly. The specific fear is that the gain came from something incidental, longer training, a better learning rate, more parameters, and that the idea the paper is named after contributed little. A single ablation table answers it, and its absence in a paper with multiple components is one of the few omissions reviewers treat as evidence rather than as an oversight.

Notation introduced after it is used, a method section that describes the implementation instead of the idea, a figure whose caption does not say what the reader is looking at.

Both big venues rate this on its own line, ICLR as presentation and NeurIPS as clarity, each out of 4. That matters more than it sounds. A number recorded separately cannot be carried by a reviewer who liked the rest, and it is the one weakness a rebuttal cannot argue away, because the thing being rated is the PDF you already submitted. It is also the cheapest of all of these to have avoided. What each NeurIPS field is for, and the same for ICLR.

Nothing is wrong with the paper. It is being judged against criteria that were never the point of the work.

This is the rejection that wastes the most time, because the paper needed no repair and the six months it costs buys no improvement. It is also the only one on this list you can settle before you submit rather than after, which is why the matcher is free and needs no account.