The peer review process at top machine learning conferences has always had its quirks, but what we are seeing at ICLR this year looks less like a quirk and more like a structural drift. A researcher who submitted applied theory work, roughly 60 to 70 percent mathematical derivation with experiments at the end, received a score of 222 on the probabilistic methods track. All three reviewers attacked the paper's novelty and asked for additional experiments. Not a single comment addressed the theory, which was the core contribution. One reviewer simply stated that he did not understand the paper or why the task existed, even though the task is well-defined with substantial prior work. This is not an isolated incident. The researcher tracked similar submissions using accept-odds and found the same pattern: reviewers ignored the math entirely and focused exclusively on experimental validation. A paper containing twenty pages of proofs received a reviewer comment that, assuming the proofs were correct (which the reviewer noted could be done by AI in 2026), the theorems only showed trivial conclusions. The rating was a 1: clear rejection. To make sense of this, it helps to look at adjacent experiences. In our piece Navigating Peer Review for Original Classical ML Research at Top Conferences, researchers described similar frustration when their mathematically grounded work was dismissed for not following the dominant experimental paradigm. And in When a reviewer flags your own prior work as a competitor, we saw how review mismatches can arise when a reviewer lacks the context to evaluate a submission fairly. The pattern is consistent: when the reviewer pool does not align with the paper's methodology, the evaluation becomes arbitrary.
Our take is straightforward: this trend is damaging the field. When reviewers cannot or will not engage with theoretical contributions, they are not just rejecting individual papers, they are signaling that entire categories of research are unwelcome. The researcher suspects that because there was no bidding process this year, many theory-heavy papers were assigned to pure computer science reviewers who lack the background to assess probabilistic methods. If that suspicion is correct, the conference has a process failure, not a quality problem. The practical consequence for authors is painful. Should they engage in rebuttal or withdraw? Given that the reviewers did not even mention the theory, rebuttal feels like arguing with people who are not reading the same paper. The odds of changing a reviewer's mind when they fundamentally do not understand the task are low.
What this means for our readers is that you need to plan your submission strategy with the current review environment in mind. If your paper leans heavily on theory, you may want to front-load your experimental section or include a clear empirical demonstration early, even if that feels like diluting your contribution. That is not ideal, but it is pragmatic. Alternatively, consider venues that explicitly value theoretical work or have track chairs who ensure reviewer-paper alignment. The researcher's decision to track similar submissions using accept-odds is a smart move, do that before you submit, not after. The specific takeaway here is that the absence of bidding this year appears to have broken the match between paper content and reviewer expertise, and until conferences address that structural issue, authors of theory-heavy work should budget for a higher rejection rate and plan their rebuttal strategy before they see the reviews. One detail to watch: if next year's ICLR reinstates bidding or introduces a track-specific reviewer pool, that will be the signal that the community has recognized the problem.