Coding agents are now sufficiently good that they can readily churn out workshop-quality papers with minimal human intervention. In my own experimentation, I've found that supplying a coding agent with a few seed papers and a generic outline of the research process (ideation, literature search, experimentation, write-up...) can produce a plausible submission after a few hours on a GPU node. More sophisticated systems (which I won't link to because I find them somewhat distasteful) incorporate researcher personas, mock reviews, etc. and have been used to produce papers that were ultimately accepted at reputable conferences (a debatable achievement, but a measurement of sophistication nevertheless). Apart from fully autonomous research, modern AI systems more broadly are likely accelerating many parts of the research process. While this effect is harder to quantify, we can see its indirect consequences in rapidly increasing submission rates—for example, NeurIPS 2026 received around 40,000 submissions (up from around 20,000 last year). Each venue is developing its own coping strategies, such as TMLR's recent decision to implement annual author submission quotas.
After the dust settles, I hope that the acceleration of the scientific process results in a newfound focus on quality rather than quantity. In the meantime, it might be prudent to brainstorm ways that we (i.e. the field of machine learning research) can change our publication venues to adapt. I've been doing this brainstorming since the ICLR 2026 shenanigans. Below are three ideas that have stuck around in my brain, roughly in decreasing order of ridiculousness.
If it requires a vanishingly small amount of human effort to churn out a mediocre paper, how can we change our publication venues so that at least some human effort is required? Research agents operate in the digital world and consequently are largely limited to producing digital artifacts. Perhaps it is time to revisit the publication cycles of yore (which, admittedly, were all before my time), which involved mailing physical pieces of paper to each other, as part of paper submission, review, and publication. To require maximal human effort, one could imagine a publication venue where all submissions (including plots and diagrams) must be hand-written. It would then be natural (though optional) to distribute the journal only in print. Of course, someone could easily outsource most of the mental and experimental effort to a research agent and then copy the result by hand, but this would be tedious. Perhaps the manual effort would provide a sufficiently high barrier that the journal would only receive submissions that had seen a significant investment of human effort and thought.
I might argue that submissions involving minimal to no human effort do not deserve to receive human review. Fortunately, in tandem with developments of research agents, there also has been a considerable amount of work on reviewer agents. Claims about the effectiveness of these agents typically center around a significant level of agreement with human reviewers (another debatable achievement, but another measurement of sophistication nevertheless). Perhaps it is most appropriate for AI-generated papers to receive AI-generated reviews. It would be relatively straightforward to set up a publication venue where all submissions went through a thorough AI review process and received an accept/reject decision in minutes. I think acceptance at this venue would not be meaningless, and could be a reasonable indication that the submission satisfies the most basic requirements of a research paper. To prevent research agents from DDoSing this venue, a modest submission fee could be charged to pay for the cost of the review. Such a fully automatic journal might stem the tide of AI-generated submissions to our established venues.
Historically, paper acceptance (and designations like spotlight, oral, or best paper) have been a noisy but somewhat meaningful indicator of potential future impact. As the quantity of research grows and our publication systems strain under the pressure, I worry that any meaningful signal begin to disappear altogether. As a result, we may only be able to get a reliable sense of a paper's impact in retrospect. On the other hand, I don't think it takes too long to get some reliable sense of the potential long-term impact of a publication—maybe only a year or so. In this light, I wonder if it would be valuable to explore a conference format that solely consists of invited presentations by researchers who, sometime in the past year or so, have published some impactful work. The program committee's job would be to nominate such researchers and vet one another's nominations. Inclusion would be via nominmation only and it would not be possible to submit one's own work in the traditional sense. Invited talks would not need to solely be about the nominated work; indeed, it probably would be most interesting if speakers presented on their ongoing and future work. I have a vague recollection that the “symposia” track of NeurIPS 2015 had a similar approach. I wonder if it is time to revive this format.
If I wanted to devote most of my time to managing research venues (I don't), I might try out all three of these ideas. But I'd be open to discussing any of them!