Where is AI currently making the most noticeable efficiency gains? Writing papers certainly ranks among the top.
This year, the number of abstract submissions to ICLR reached an astonishing 60,000+, surpassing the total number of submissions from all years since ICLR's inception in 2013 through 2026.
More specifically, the first ICLR received 67 submissions, while ICLR 2026 saw this number rise to 19,525. When summing up all 14 conferences from 2013 to 2026, the total reaches approximately 56,000 papers.
Now, another ICLR 2027 has come and gone.
The contributors, upon seeing this number, would likely be collectively stunned and collapse in shock, as if witnessing an atomic explosion.
Some netizens joked that in a few years, the number of submissions could reach the millions...
An internet user did the math: even with a 20% acceptance rate, 60,000 submissions would equate to 12,000 papers.
What does it mean to have 12,000 papers recorded at a conference?
The previous ICLR 2026 accepted 5,355 papers. NeurIPS 2025 accepted 5,290 papers, CVPR 2025 accepted 2,872 papers, and ICML 2025 accepted 3,260 papers.
Based on this purely mathematical scenario, the number of accepted papers at ICLR 2027 would be equivalent to two editions of NeurIPS 2025, four editions of CVPR 2025, or approximately four editions of ICML 2025.
Of course, the "60,000+" currently discussed in the community more accurately refers to the submission volume that emerged after the ICLR 2027 abstract deadline.
The abstract submission deadline for ICLR this year is September 18, and the full paper deadline is September 25. The official guidelines require abstracts to be genuine and valid; duplicate or placeholder abstracts will be removed. Subsequent steps include full paper submission, withdrawal, and formatting checks, so the final number of papers entering the official review process may still change.
But the figure of 60,000 is already absurd, and the resulting review pressure is obvious.
Yesterday, we reported on JEV, the recently popular ultra-efficient judgment model, and some readers joked in the comments, "ICLR reviews are saved."
One of Jev's recent notable applications is serving as a Judge for Agents: given a set of execution results, the model rapidly determines whether the task has been completed, whether the outcomes meet requirements, and what the next steps should be. Now, a previously unprecedented Judge problem has emerged before ICLR.
Sixty thousand papers—who will judge them? This has become a shared challenge for top AI conferences over the past few years.
ICLR 2026 received a total of 19,525 valid submissions, of which 779 were desk-rejected and 5,042 were subsequently withdrawn, leaving 13,763 papers to proceed to the acceptance or rejection decision stage. To handle these papers, 18,054 reviewers submitted a total of 76,139 reviews.
In other words, when the number of submissions reaches the tens of thousands, a review network of nearly 20,000 reviewers is already required. Now, the number of entries is suddenly heading toward 60,000.
ICLR itself clearly sensed the pressure in advance. In early September this year, the ICLR 2027 Program Chairs published a dedicated article introducing the new submission policy, emphasizing one key term: capacity control.
This year, each author may appear in a maximum of 20 submissions. For papers where all authors lack qualified reciprocal reviewer eligibility, each author may participate in only one such paper. Additionally, authors appearing in three or more submissions must review at least six papers;原则上, each submission must also have at least one author registered as a reviewer.
https://blog.iclr.cc/2026/09/02/submission-policies-for-iclr-2027/
The background provided by ICLR is also interesting.
The Program Chairs explicitly stated in their policy statement that competition for AI and industry positions is driving researchers to increase their paper output, and AI tools are further accelerating this trend. They also noted that some conferences have observed a significant increase in submissions from author groups lacking qualified reviewers; in ICLR 2026, approximately 20% of submissions had no reciprocal reviewer.
This year’s Call for Papers includes a special reminder for authors submitting to ICLR for the first time.
The organizers state that better AI tools are expanding the researcher community; advancements in Coding Assistants, theoretical research, improvements in research infrastructure, and the spread of machine learning into more disciplines are all driving rapid growth in research output. At the same time, the organizers encourage researchers to channel the time saved into more comprehensive and ambitious research projects.
Even pre-submission checks are beginning to be AI-powered.
On September 10, ICLR officially announced a collaboration with Google to make the Paper Assistant Tool (PAT) available to submitters. Authors can submit their papers to the system prior to formal submission to receive automated feedback. The official statement emphasized that this feedback is private to the authors and will not be included in the formal review process.
https://blog.iclr.cc/2026/09/10/making-googles-paper-assistant-tool-pat-available-to-iclr-submitters/
In a sense, this year’s ICLR itself was like a large-scale experiment in AI research productivity.
Over the past two years, large models have fully integrated into the academic paper production pipeline—helping with topic selection, coding, running experiments, polishing papers, and drafting rebuttals, while AI-assisted reviewing has also begun appearing on the reviewer side.
People are even complaining about the AI feedback loop: AI writes the paper, AI reviews the paper, and the author then has AI write the rebuttal.
Papers are increasing in number, peer review is becoming more competitive, and complaints in the community about “top conferences becoming increasingly watered down” are also growing.
Once the 60,000+ figure was released, everyone began offering advice to ICLR.
Some have suggested canceling ICLR 2027 entirely and redesigning the system to restore the original prestige of top conferences.
Others have proposed more specific solutions: each author may submit a maximum of 3–4 papers; condense the main text to 3–4 pages to ensure key results appear early; publicly release code, data, and intermediate results alongside submission for Agent verification of reproducibility; and finally, provide reviewers with tangible monetary compensation or genuine academic credit.
Someone else added: $10 is charged for each submission, and it’s non-refundable. With 60,000 submissions, that amounts to $600,000—enough to cover at least the reviewers’ coffee expenses.
Did you submit to ICLR this year? What do you think about this explosion of submissions?
This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), edited by Ava