Anger at ‘arbitrary’ conference rejections after submissions soar

Leading academic conferences have been accused of wielding a “guillotine” over research papers after a record number of submissions forced organisers to cut the number sent for peer review.

Last month, the number of abstracts submitted to the International Conference on Learning Representations (ICLR) 2027 surpassed 62,000 – triple the 19,500 submitted the year before.

ICLR is one of the top three machine learning conferences. The events are often preferred to journals in the field because of their quicker reviewing turnaround. Submissions to the other two prestigious conferences, NeurIPS and ICML, also rose sharply from 22,000 to 31,000 between 2025 and 2026 and from 12,000 to 24,000 respectively.

This year, ICLR 2027 organisers introduced a desk rejection policy whereby a proportion of papers were immediately declined, as a way of reducing the burden on its open peer review system. But scholars on the receiving end of this change have criticised the move for being too harsh.

Chilean mathematician Mircea Petrache, who had his submission rejected, told Times Higher Education: “It felt like the conference was overwhelmed with submissions and a guillotine was put on to cut down the numbers.”

Petrache, an assistant professor at the Pontifical Catholic University of Chile, said it was particularly concerning because the paper’s lead author was a PhD student.

“It’s so important for them [PhD students] to get the visibility and the contacts that happen when you go to these big conferences,” he said.

The concerns were echoed on social media, where one user said: “A colleague of mine was rejected simply because another person with the same name had also submitted a paper to the same conference.” They called the debacle a “serious issue”.

Responding to complaints, the ICLR organisers told THE that “some mechanism to filter for gross lack of clarity and content was necessary due to the influx of poorly written AI-generated papers”. They added that “no paper was rejected without at least two pairs of experienced researcher eyes both approving.”

The criteria given to reviewers was to flag “papers that are clearly of low quality or clarity”. Desk rejections are commonly used in journals and have been recently introduced in the Journal of Machine Learning Research, one of the field’s leading publications.

On the reviewing side, academics also had problems. Machine learning scholar Bastian Grossenbacher-Rieck told THE that the field is “suffering from irresponsible actors as much as the rest of the system”.

As a senior reviewer for ICLR, Grossenbacher-Rieck, professor of machine learning at the University of Fribourg, said he had rejected papers on the grounds that they made “incomprehensive claims” that were not backed up by the text. But, in his view, the bar for desk rejection was “relatively high”.

Although he admitted that “I know – even from my own colleagues – that there were some desk rejects that I would find on the face of it a bit questionable.” In those “unlucky” cases, “I fully empathise” with those people, he said.

The organisers said they “acknowledge that a desk rejection might be disheartening, particularly to newcomers to the field, but we also hope that the standards we introduce percolate through the culture and raise the publishing standards throughout the community”.

In these “fast-paced times”, they added, “it is quite likely that many conferences will experiment with different solutions”. The community will learn “from trial-and-error and come out on the other side with improved mechanisms”, they added.

They said it is also likely that “filtering mechanisms – whether they’re done solely done by humans or by a combination of humans and AI – will have to be in place in the near future”. This will improve as the understanding of which tasks AI systems are “robustly good” at grows.

In future, “we will hopefully be able to set up systems in which human judgement is robustly the deciding factor, while enough sanity and correctness checks are being performed by AI to alleviate the load on the community”, they added.