Same problem, different field

Pairs of papers from different domains whose plain descriptions say the same thing. Volunteers' AI models describe each paper without its field's words; when two papers far apart are described alike, they may be working on the same problem without knowing it.

Each pair is a lead for a reader to judge, not a finding. Only papers that two different models described consistently are paired, so a vague description cannot make a pair. How it is computed is at the end of this page.

Pairs found so far

Models, compared

How closely each model's descriptions agree with other models' descriptions of the same papers, next to how alike descriptions of two different papers are by chance. Agreement is a check that a model says something specific about each paper; it is not proof that what it says is right.

How it is computed

• Each paper is described by two volunteers, each with the model of their choice, as the problem it works on and how, without the paper's distinctive words. The descriptions and the model that wrote each are on the data branch, under CC0.
• Descriptions are compared by the cosine between their TF-IDF word vectors. It is simple and anyone can recompute it; whether a learned embedding finds better pairs is a question for the data once there is enough of it.
• Chance is how alike descriptions of two different papers are, from a fixed random sample. A paper is used only if descriptions of it by two different models agree more than 95% of those chance pairs do. Two volunteers running the same model do not count as agreeing.
• Pairs within the same domain (life, health, physical or social sciences) are left out, and no paper appears in more than two pairs. The code is pipeline/combine.py.