mlsanityData to 7 Oct 2026

About

15 areas · 193 directions · 97,621 papers on the map. Drawn by machine from arXiv and reference lists; not checked by hand. Treat it as a best guess.

How it’s made

  • Papers that cite the same work cluster into directions.
  • Reference lists known for 83,265 of 97,621 papers.

Confidence

Confidence is about the grouping, not the name: how clearly a direction’s papers form one topic of their own. It averages three checks: how close the papers are in content, how many belong here rather than in a neighbouring direction, and whether the group comes back when the area is re-clustered.

High is a score of 0.55 or more, medium 0.35 to 0.55, low below 0.35. Low can mean a mixed topic, or a clear topic that overlaps a neighbour.

We checked the levels against a model judge and 20 hand checks. About 1 in 7 directions change level when an area is re-clustered.

Placement

Each new paper is placed in one direction. Some carry a note:

also fits: direction
A near tie: the paper fits that direction almost as well, and may belong there. The link goes there.
no mark
Placed here; no other direction came close.

Not on the map

A paper joins an area by its own category, key phrases and wording, not by who cites it. So some much-cited papers sit outside. They still have a page and show up in search.

Names

Names come from a matching expert-survey section where there is one; the others are short descriptive names, checked by hand against each direction’s papers.

Lineage

The papers a direction builds on, oldest first. The red line is the longest chain: each paper cites the one before, so each is built on by the next.

Earlier versions drew lineage as this table: each row is a paper, each filled cell a paper it cites, red the longest chain. The logo is its top corner. Embodied and robot world models.

Sources

Data: arXiv · Semantic Scholar. Rebuilt daily; through 7 Oct 2026. Fonts: EB Garamond, Inter Tight (OFL).