brontide.
a research atlas
Unfolding the landscape…A moment for the details.

Ideas, in motion.

2.7M papers. 64 fields. Two decades.

VERTICAL AXIS · PUBLICATION YEARDrag to orbit · scroll to zoom
2007 — 2026publication years
Points: annual · statistics: monthlyLoading sample
A FIELD GUIDE

A little context.

Real papers, a fixed map, and a few useful limits.

Reading the currents

Each ribbon is one territory. At each month, its vertical thickness is the territory’s paper count over the trailing 12 months divided by the corpus count over the same window. These are count-weighted shares, not an average of percentages. All 64 ribbons sum to 100%. Early 2007 uses the available 1–11 months; no earlier observations are invented. The 234 observations run from January 2007 through June 2026. The final year is partial.

The territory order is fixed by ID. A data-derived wiggle baseline minimizes visual movement; it is only a layout choice. Position and color do not encode transfers, citations, or influence between fields. Adjacent trailing-window values are joined by a smooth, nonnegative interpolation. Hover, focus, or tap for the exact window and its count-weighted share, alongside the raw monthly count and share. Search gives access to even the thinnest ribbons.

The year sliders highlight a window while keeping the full chronology in view. The playhead marks the final available month of the selected year. Currents uses observations only; the separate territory chart contains the archived forecast experiment.

What the axes mean

Horizontal position is a fixed PCA → UMAP projection of SPECTER2 paper embeddings. Nearby dots often describe related work, but projected distances and apparent gaps are not calibrated semantic measurements. Clustering uses the 50-dimensional PCA space.

In space-time, height is publication year. In density, height is a smoothed, normalized concentration of the sampled points in the selected window. Density is neither paper quality nor a forecast.

What you are seeing

The map displays a stratified sample of the frozen arXiv corpus. Counts and monthly histories use aggregate artifacts, not dot counts.

Map points have annual dates; charts retain monthly resolution. The final year is partial. Topic names are generated descriptions, not an authoritative taxonomy. Representative titles may come from outside the selected window.

Microregion series can be inspected within a territory. The public point sample has only superregion assignments, so microregion selection does not relabel or relocate dots.

Window and comparison

The sliders filter publication years. Compare retains papers before the selected window as a muted baseline and highlights papers in the window. The chart marks the same window; percentage-point changes compare mean observed monthly shares in the endpoint years.

This is retrospective exploration: the display projection was fitted using the full frozen corpus. Moving the year does not create a historically refitted model.

Forecasts and evaluation

Forecasts begin after June 2026 and use the committed production model, SmoothedMomentum. They are independent projections, not jointly normalized across regions, and are experimental extrapolations rather than promises. The frozen selection procedure beat its momentum baseline at 0 of 3 masked cutoffs, missing its ≥2 criterion.

Published 80% intervals failed the calibration criterion; the explorer does not present them as calibrated uncertainty. The frozen result and its scope remain in the repository.

Read the data guide
Make it yoursCurrents: ↑/↓ fields, ←/→ months, Enter details. Maps: drag to orbit, Shift/right-drag to pan, scroll/pinch to zoom, arrows to orbit, +/− to zoom, Home to reset. / search · Escape clear