What’s next for machine learning peer review?
A bit over a year ago, I wrote about the dangers of using LLMs for peer review. The most serious concern I had was algorithmic monoculture: the research community would collectively end up optimizing to meet the taste of a few LLMs (or LLMs with correlated preferences), which seems very undesirable to me.
These days, peer review in ML is in an even more clearly untenable position. There are more papers every year, the reviewing system can’t keep up, many reviews are poor-quality (or clearly substantially LLM-generated). As an area chair at NeurIPS, I find that the majority of reviews are not particularly helpful to me in deciding whether a paper should be accepted (though I’m very grateful to the smaller number of reviewers who provide highly informative assessments). This is not a unique experience; clearly no one is happy with the current system.
LLMs are also much better than they used to be. Does this increase in capabilities solve the potential problems of relying on them for peer review? On the one hand, I think LLMs are clearly capable now of assessing the soundness of papers at a level that exceeds the average quality of the current review process: they are better at catching errors in proofs, mistakes in code, deviations from standard experimental practice, and so on. They are more likely to know about related work, and less likely to misunderstand what a paper is about. With some effort at writing clear specifications of what a given research community expects in terms of methodological soundness, I would expect current frontier models to be significantly better at applying such specifications than either the average ML conference reviewer or any practically implementable improvement to the current system. To be clear, this is an area where investing in good evaluations would be valuable; I’m not advocating making a decision purely based on vibes. I’m also not claiming that LLMs are currently better than the best humans at these tasks, but beating the current system is a pretty low bar.
The concern about enforcing a monoculture in taste, however, is not solved by increases in capabilities. One thing I like about the current review process is that its sheer randomness means that starkly different ideas, even if disadvantaged, have some real probability of getting through. From talking to people in other fields where the review system is more closely controlled by high-status researchers (e.g., economics) I get the sense that researchers in ML actually have a bit more freedom in some ways to pursue their own taste – there’s less point in optimizing intensely for outcomes from a highly random process. I wouldn’t want to sacrifice the rapid cycle speed and relative freedom that computer scientists enjoy just to get a lower-variance review process; doing so would reduce the overall ability of the field to make progress.
What to do? My proposal is that we build a process with two components. First, LLM-generated certificates of soundness. These could be associated with particular research communities or professional organizations. But, beyond soundness, the other function of the peer review process is to direct attention: among all of the (many) papers that are sufficiently methodologically sound, which ones are important to researchers working in a particular area?
The second component would be an ecosystem of LLM-curated recommendation systems for papers. Previous recommendation systems were limited to recognizing general subject areas and the like, but LLMs can render much more precise judgments about the intellectual content of a paper. I think it’s likely that, with some investment, potentially including finetuning, a given researcher could teach an LLM near the current frontier to do a pretty good job of recognizing papers that will be to their individual taste. Again, not perfect, but certainly better than scanning through the NeurIPS proceedings. If I built such a thing, other people could subscribe to get a feed of “papers that Bryan’s algorithmic proxy recommends”; perhaps if you think I have good taste, you would sign up for such a feed. The important feature of this system is to deliberately produce a variety of models that implement the preferences of individual researchers or groups who have coherent views on what research is important. The research community can then vote with their feet in which to pay attention to. Individuals should get recognition for producing algorithmic feeds with wide followings, or which identify papers that are later broadly recognized as impactful – it’s a direct sense in which they have good insight into what work is important.
Since LLMs scale far better than human attention, I believe that such a system would increase the chances that important work by less-famous authors is recognized, compared to either the current conference system (NeurIPS or ICML is too large, and the review process too low-signal, to direct significant visibility to most accepted papers) or its current practical alternative (trying to go viral on X, which is effectively relying on a bad recommendation system instead of one deliberately designed for the research community).