We Made Research Faster. Why Are We Surprised It's Moving Faster?
We need to redesign peer review for abundance.
I keep seeing people shocked by the explosion in submissions to conferences like ICLR.
But shouldn’t we have expected this?
First, the scale is worth putting in perspective: ICLR 2026 received 19,525 valid submissions; for ICLR 2027, early reports indicate that more than 60,000 abstracts have already been submitted - over 3× last year’s total. The final 2027 paper count may be lower, since the full-paper deadline is September 25.
Still, the growth is remarkable!
ICLR received 4,938 submissions in 2023. In 2025, that number reached 11,603. In 2026: 19,525.
That’s nearly a 4× increase in three years.
But I don’t think this is an anomaly.
It is the continuation and acceleration of something that has been happening in science for a long time: technology keeps reducing the cost of turning an idea into an experiment, and an experiment into a publication.
- Better programming languages did it.
- Open-source software did it.
- HPC and Cloud computing did it.
- GitHub, arXiv, GPUs, large datasets, and global collaboration did it.
And now AI is doing it at a completely different speed.
The broader numbers tell the same story. According to Stanford’s AI Index, the number of AI publications grew from roughly 102,000 in 2013 to more than 242,000 in 2023.
Generative AI is simply pushing the marginal cost of research iteration even lower.
You can prototype faster. Write code faster. Run experiments faster. Analyze results faster. Search literature faster. Draft faster. Revise faster.
So, naturally, we are going to submit faster too.
And yes - bad ideas will become faster as well.
That part matters.
Reducing the cost of experimentation doesn’t magically improve human judgment. It accelerates brilliant ideas, mediocre ideas, incremental ideas, and terrible ideas alike.
But volume alone doesn’t tell us that the ratio of good ideas to bad ideas has suddenly collapsed.
What it tells us is that our research-production pipeline has scaled faster than our research-evaluation pipeline.
And that is the real problem.
We cannot have AI-assisted research operating at 2026 speed while peer review operates with essentially the same human workflow we designed for a much smaller research ecosystem.
The answer cannot simply be:
“Researchers should produce less.”
We need to redesign peer review for abundance.
Use AI for reviewer-paper matching, semantic clustering, duplicate and citation checks, methodological consistency checks, reproducibility assistance, review-quality checks, and identifying papers that deserve deeper human attention.
Not to replace scientific judgment.
To help human judgment scale.
Interestingly, this is already beginning. At ICLR 2025, an experiment provided LLM-generated feedback to reviewers. 26.6% of reviewers who received that feedback revised their reviews, incorporating more than 12,000 individual suggestions from the system.
That is the direction I find much more interesting than debating whether researchers are “writing too many papers.”
We built tools that dramatically increase intellectual and engineering throughput.
Now people are using them.
The productive question is no longer:
“How do we stop people from producing so much?”
It is:
“How do we build systems capable of finding the exceptional work in a world where producing work is becoming dramatically cheaper?”
Every major technological shift creates abundance before institutions learn how to manage that abundance.
AI research is experiencing that transition right now.
We built acceleration.
Now we need to build the infrastructure that can keep up with it.