What the papers at NeurIPS, ICML, ICLR, CVPR, ICCV and ECCV are about, read from their titles and abstracts. Shares, how they changed, and which papers combine which topics.
2026 so far: ICLR, ICML, CVPR and ECCV. NeurIPS 2026 meets in December; ICCV meets in odd years and ECCV in even years.
Share = percentage of papers whose title or abstract names the topic; Papers = the number of those papers. A paper can count for several topics. The first row gives all papers in the data. "Change" is first to last year in percentage points of share; "per year" is the change in the topic's odds per year. Trend is called only when significant after correction for testing all topics (q < 0.05). Hover a number for its 95% interval and paper counts. Click a row to see its papers.
Each chart is one topic, by year. Papers: bars, the number of papers whose title or abstract names the topic. Share %: line, the same share as in the Topics table, with its 95% interval as a band. Each chart has its own scale unless "Same scale" is ticked. Hover a year for its numbers; click a chart to see its papers.
Share of each topic in each venue's latest year (2026 for ICLR, ICML, CVPR and ECCV; 2025 for NeurIPS and ICCV). ↑ / ↓ mark a trend over the whole period that is significant within that venue. Darker cells are larger shares. Click a cell to see those papers.
Word size = number of papers. Click a topic to keep only its papers; click more topics to keep the papers that have all of them; click a selected topic again to remove it.
Main-track papers of NeurIPS, ICML and ICLR (machine learning) and CVPR, ICCV and ECCV (computer vision), 2020 to October 2026, with an abstract (–). Workshops, co-located events, Tiny Papers, position papers and competition or demo tracks are left out. The list of papers comes from the dblp XML dump of 2026-10-13, plus ICML 2026 from its PMLR volume (the dump did not have it yet); position papers are left out by their "Position:" title prefix as well. Abstracts come from each conference's own pages wherever possible: PMLR (ICML), the NeurIPS proceedings, the ICLR proceedings (2024–2026) and the ICLR virtual site (2020–2023), CVF open access (CVPR, ICCV) and ECVA (ECCV); OpenAlex by DOI for CVPR and ICCV; when dblp writes a title slightly differently from the conference's own listing (formulas, typos, a method name added), the most similar title of the same venue and year (145 pairs, each read); the arXiv version, only when the title is identical, for what remains. OpenReview's API refuses anonymous requests and was not used. Code and project links are removed from abstracts before matching.
Topics were drawn from the most frequent and the fastest-changing title phrases of the six venues, then written as rules (keywords and phrases). A paper counts for a topic when the rule matches its title or its abstract. Words that are generic inside abstracts (dataset, attention, reasoning, agents, fine-tuning, optimization, distillation, unsupervised, face) have stricter abstract rules. For every topic, random papers matched only through the abstract were read, in 2020–2023 and in 2024–2026, to check that the words keep the topic's sense; rules were corrected until the samples held up. The method is the same as in the robotics study.
Shares are standardised to the venue mix of 2024 and 2025 together, so that both ICCV and ECCV are in it and the years compare. A venue that did not meet in a year (ICCV in even years, ECCV in odd years, NeurIPS in 2026) contributes its share from the model, with a 95% interval. The trend is a logistic regression of topic on venue and year over the whole period; a topic is called rising or declining only when significant after Benjamini–Hochberg correction across all topics. "2024 onwards" repeats the test on 2024–2026 alone. Per venue, the trend is a likelihood-ratio test.
Scripts, topic rules and result tables are kept with the study. Abstract texts are not redistributed; the scripts fetch them from the original sources.