Scores for ICLR 2026 are out, reviewers angrily call it "crazy", and DeepMind researchers teach you how to make a comeback from a desperate situation.

The review results of ICLR 2026 are out, shocking the academic community: the number of submissions has skyrocketed to nearly 20,000, but the scores have taken a nosedive. The average score dropped from 5.12 to 4.2. Reviewers are complaining about the poor quality of the papers, even suspecting that some are AI-generated. Why has this academic feast lost its flavor?

The review results of ICLR 2026 have been announced!

The number of paper submissions for this year's ICLR has reached a record high.

According to the statistics on the Paper Copilot website, compared with the total of 11,672 submitted papers in 2025, there were 19,631 submissions for ICLR 2026. Although the quantity increased, the quality declined -

The highest score in ICLR 2025 was a perfect 10, while in ICLR 2026, it was 8.5;

The average score in ICLR 2025 was 5.12, and in ICLR 2026, it was 4.20.

Link: https://papercopilot.com/statistics/iclr-statistics/iclr-2026-statistics/

Many submitters have experienced getting a score of 0 on ICLR for the first time in their lives:

Some domestic submitters even had all three of their submissions rejected. One of them tweeted, exclaiming about this unique life experience:

ICLR 2026 will be held in Rio de Janeiro, Brazil, from April 23rd to 27th next year.

Like NeurIPS and ICML, ICLR is one of the three most influential conferences in the field of machine learning and artificial intelligence research.

ICLR, short for "International Conference on Learning Representations", was founded in 2012 by Turing Award winners Yann LeCun and Yoshua Bengio.

The average score of ICLR 2026 is nearly 1 point lower

Overall, the average score of ICLR has dropped by nearly one point compared to last year:

In 2026: The top 30% of submissions scored over about 5.0 points.

In 2025: The top 30% of submissions scored over 6.0 points.

Afshin Khadangi, a biomedical engineering doctor, used an API to obtain 19,129 review comments and summarized some of the highlights:

Only 1,792 papers (about 9%) had an average score of 6 or above.

One paper (id 12681) received 9 comments!

In individual scores, the average was 4.22, with a standard deviation of 1.86, based on 74,371 comments.

Reviewer Cheems Wang noticed that in his field, papers always had at least one extreme score, and the scoring this year was indeed more conservative.

Reviewer Yu Su also noticed that the average score was quite low. Only 18 out of 225 papers had an average initial score of over 6, barely meeting the conference's acceptance threshold.

A doctoral student from Tsinghua University and a postdoctoral fellow at Princeton University discovered an obvious scoring pattern in ICLR 2026:

The higher the submission ID, the lower the score seems to be.

Someone noticed a similar pattern two years ago:

Surprisingly, there was a "psychotic" comment at the review site:

This intense review scene attracted the attention of Reddit users:

Now the original review comment has been deleted. The reviewer apologized for the offensive remarks but insisted that there was no mistake in the problems pointed out in the paper.

Another ICLR reviewer from Amazon also had complaints:

Right after the review comments were announced, one of the papers I was responsible for was withdrawn.

Moreover, he has encountered similar situations several times at multiple top - tier conferences this year.