Can authors rank their own papers better than peer reviewers?

At ICML 2023, the paper's team asked researchers with several submissions to rank them by perceived quality, through a platform they built (OpenRank.cc) with the conference organizers' approval. The paper reports that 1,342 researchers ranked 2,592 submissions, and that the papers each author ranked highest drew, on average, twice the citations of the ones ranked lowest, for accepted and rejected papers alike. Of the 22 papers with over 150 citations, 17 had been ranked highest by their authors. The sharpest comparison is that self-rankings "outperformed peer review scores in predicting future citation counts."

The paper grounds this in who knows the work best. Authors "possess unique understanding of their work's conceptual depth and long-term promise," while overloaded reviewers "may prioritize quantifiable gains over a submission's broader scientific implications." The comparative format matters because authors "cannot simply claim all their submissions are of the highest quality." Because each author's highest and lowest papers are compared, the design also roughly matches on author background and research area. The paper cites game-theoretic work (Su, 2021, 2025; Yan et al., 2025) for the claim that this format can incentivize truthful reporting under certain conditions.

Set against the nearest notes, this asks a different question from Does banning LLM use in peer review change review outcomes?, which measures whether review policies move scores. This paper asks which signal review should draw on at all. It also gives empirical footing to the warning in How much does rhetorical style shift AI review scores?. The discussion says LLM use "may further bias evaluations toward surface-level characteristics rather than profound scientific contributions," and the rhetoric study shows LLM scores moving with presentation. The contrast with Can inference scaling help reviewers catch errors humans miss? is sharper. That paper searches manuscripts for correctness failures, while this one separates "methodological correctness" from "likely scientific consequences." A reviewer tuned to the first may not capture the second.

The excerpt leaves much unestablished. Citations are the paper's own proxy for impact, which it calls "widely used, albeit imperfect." The evidence comes from one conference and one survey year. The paper says similar experiments at ICML 2024 and 2025 and NeurIPS 2025 have already been conducted, but it reports none of their results. The citation analysis covers 1,527 submissions, not 2,592, after keeping only exact title-and-author matches on Semantic Scholar, and the survey response rate was 30.4%. The excerpt gives the preprint check (mean posting dates of March 4 and March 8, 2023) but cuts off before the self-citation result, and it reports the Google Scholar and GitHub-star results without figures. The defensible implication is that a self-ranking step is worth testing as a complement to review scores for finding influential AI work. The excerpt does not show that self-rankings should replace review, or that they track quality beyond citations.