128,400 papers · 19 venues · 2013–2026 · topics discovered from the search index itself
Topics are discovered by clustering every paper's embedding — no keywords, no manual tagging. Pick which venues to include and over what period; everything below recomputes instantly.
The busiest research topics, ranked by papers published in a trailing window, reordering as the field changes. Each bar is split by conference, so you can watch a topic that starts in one community pick up colours as it spreads. Topic names are tinted by momentum — warm for rising, cool for fading — and the arrow shows how many places it has moved in two years.
Each band is a broad research theme; height is its share across the selected venues, counting each venue's most recent meeting as of that date. Bands sum to 100%, so this is the changing composition of the field — growth in one area necessarily comes at another's expense. The line steps whenever a conference actually meets. Click a band — or its legend entry — to break it into its constituent topics.
Change in share of the literature between the first and last three years of the selected window. Hover any bar for representative papers.
Each line is one venue's share devoted to the selected topic, placed at the month that venue actually met. Some topics are born in one community and diffuse outward; others never leave home.
Every topic as a row, normalised to its own peak and sorted by when that peak occurred. Older topics burn out toward the top; newer ones light up at the bottom-right.
Showing the 32 largest topics — all 100 are shown on a wider screen.
Every paper in the index is embedded as a 1024-dimensional vector. Those vectors are clustered into 100 fine-grained topics with spherical k-means, which are then grouped into 14 broader themes by Ward agglomerative clustering. Topic names are hand-written from each cluster's distinguishing terms and its most central papers. No paper is counted twice: each belongs to exactly one topic.
Our coverage of each venue is not constant over time. ICLR, for instance, has fewer than 20 papers per year indexed before 2020 and over 800 afterwards; SIGIR and TheWebConference only begin in 2024; KDD only in 2026. If you compare a year when a venue was absent against a year when it was present, you will see a dramatic "new topic" that is really just that venue arriving. The default balanced panel avoids this by using only venues with complete, unbroken coverage. The warning above fires whenever your current selection reintroduces the problem.