An open network that looks for research in one field that another field could use but has never cited, and measures how often it is right.
In 1986 the information scientist Don Swanson read two sets of papers that had never cited each other. One showed that dietary fish oil lowers blood viscosity, reduces platelet aggregation and dampens vascular reactivity. The other described patients with Raynaud's syndrome, whose condition involves high blood viscosity, platelet aggregation and vascular reactivity. Swanson proposed fish oil as a treatment. Three years later a small double-blind trial (32 patients; DiGiacomo, Kremer & Shah, American Journal of Medicine, 1989) found that it helped patients with primary Raynaud's, though not those whose Raynaud's came from another disease. He called this undiscovered public knowledge: results that are already published, but in literatures that nobody reads together.
Swanson worked by hand, in medicine. This project tries to do the same thing for every field of science, openly, and checks its own predictions against what actually happened.
First result, October 2026: inconclusive. We froze the literature at the end of 2017 and ranked 28,114 pairs of topics from different domains that had barely cited each other. 305 of them (1.1%) started citing each other by 2023. A model of the citation network around each pair ranked them well: about 14 times better than chance, and 30 of its top 100 pairs did connect. Adding what the papers say, measured by the words they share, did not measurably improve it (difference in average precision -0.010, 95% interval -0.050 to +0.045). Under the plan written before the analysis that is not enough to take the method further as it is. On a second, independent sample of 300 topics the network ranking did as well (about 16 times better than chance) and adding the words was again inconclusive. All the numbers, the robustness checks and the code are on the results page. Rather than spend weeks of data on a bigger version of the same comparison, the project now tests the stronger idea: describing papers without their field's jargon, which volunteers do with their own AI models (step 4 below).
- A map of who cites whom. OpenAlex, a free and open index of scholarly works, assigns about nine in ten works to research topics. For each topic we take the 100 papers most cited up to the freeze, and record which other topics cite them and in which years. Volunteers collect this map on their own computers, with their own free OpenAlex allowance (client); this job needs no AI model.
- Candidate pairs: two topics from different domains (physical, life, health or social sciences) whose papers have barely cited each other's most cited papers yet. Each pair is scored on the shape of the citation network around it. Adding how alike the two topics' abstracts are, word for word, did not help in the first test.
- A backtest anyone can rerun. We freeze the literature at the end of 2017, rank the unconnected pairs, and check which of them did start citing each other between 2018 and 2023. The ranking model learns its weights on an earlier freeze (2011, with outcomes from 2012 to 2017). One input still comes from today's OpenAlex: the topic each paper is filed under. The plan says how that is checked. A forward test, with no hindsight at all, was registered on 12 October 2026: the same model ranked pairs at the end of 2025, and the ranking, published with a timestamp before its outcomes exist, is scored on citations from 2026 to 2031 (forward/).
- The problem, without the jargon. Fields often name the same problem differently: estimating a hidden state from noisy measurements is a Kalman filter to a control engineer and data assimilation to a weather forecaster (the example is Kulikowski's). A language model can rewrite a paper as the problem it solves, without its field's vocabulary. In a recent preprint this raised the average precision of matching papers on the same problem across fields from 0.22 to 0.51, on 109 papers (Kulikowski, 2026). That is one model call per paper: too many for one project, but manageable when many volunteers each run their own model, local or free in the cloud, and that is the work volunteers do now (take part). The descriptions go into an open index, and papers far apart that are described alike are shown as pairs (same problem, different field). Any model good enough to write them has read papers from after any past freeze, so whether they predict connections is tested forward: described now, judged on the citations that arrive later (draft).
- Experts and checks. Researchers rate the proposed connections that touch their own field, because a connection that will happen is not always one worth making (see SciMuse below). Every description and verdict records which model produced it. Agreement between different models, measured against chance, and, next, pairs with known answers mixed into the work measure how far to trust each one.
It costs the project nothing: the data is open, the site and the coordination run on GitHub, and the work runs on contributors' own machines. Contributors who choose a paid model pay for their own calls.
Now:
- Take part: lend your computer and an AI model of your choice, on your own machine or free in the cloud, to describe papers in plain words. The page shows the progress so far, who has helped and which models they used; the pairs of papers the descriptions bring together, and how the models compare, are on Same problem, different field.
- Find a flaw in the first test's results or the forward test's plan, and tell us what could make a result wrong. Criticism of a plan is worth most before its data is collected, while it can still change in the open.
As each part ships, you will be able to:
- rate the proposed connections that touch your field;
- run the tool on your own machine with a paper, an ORCID iD or a short description, and see which distant fields work on the same problem. Nothing leaves your machine unless you choose to share it.
Clients follow one protocol. Two collect the map now: the Python client, and the Aether client, a single program written in Aether that needs no Python. On the same unit they write the same file, byte for byte. How the pieces fit together, and in what order they are built, is in DESIGN.md.
The idea is not new. This project builds on the work below. Its backtest has to beat a baseline built on the kinds of network features that did best in Science4Cast, including how fast topics and their neighbourhoods grow, though in a simpler model than the best entries used.
- Literature-based discovery: Swanson's method and his ARROWSMITH software, and the time-sliced evaluation that later became a common test.
- mat2vec (Tshitoyan et al., Nature, 2019). Word embeddings trained on materials-science abstracts published before 2009 identified thermoelectric materials that were only reported years later.
- Bridger (Portenoy et al., CHI 2022). Describes authors by the problems and methods in their papers, and suggests authors who share some of them but sit outside a researcher's usual circles. In user studies with computer scientists, its suggestions were judged more interesting and novel than those of a relevance-focused baseline.
- Science4Cast (Krenn et al., Nature Machine Intelligence, 2023). A benchmark for predicting which concepts will be studied together, built from more than 143,000 AI papers. Its finding that carefully chosen network features beat end-to-end learning is why those features are our baseline.
- Human-aware AI (Sourati & Evans, Nature Human Behaviour, 2023). Modelling which scientists could plausibly make a discovery improved the prediction of future discoveries by up to 400%.
- SciMuse (Gu & Krenn, 2024, revised 2026). A knowledge graph built from 58 million papers and GPT-4 generated personalised research ideas. More than 100 research group leaders at the Max Planck Society scored over 4,400 of them: the mean was 2.40 out of 5, and 24.9% received 4 or 5. Choosing concept pairs with the knowledge graph did not make ideas more interesting than giving the model paper titles alone, and pairs predicted to be highly cited leaned slightly less interesting. Predicting a connection is not the same as finding one worth making, which is why expert ratings are part of this project.
- Idea-Catalyst (Kargupta et al., 2026). Restates a research goal as problems free of any one field's terms and looks for how other disciplines have approached them. It reports gains of 21% in novelty and 16% in insightfulness.
- Same Problem, Different Field (Kulikowski, 2026, preprint). Rewrites each paper, with one language-model call, as the computation it performs without its field's vocabulary. On 109 papers from 18 method families this raised cross-field retrieval average precision from 0.222 to 0.513, while four trained scientific embedding models did worse than simple word matching (TF-IDF) on the abstracts: they capture topic and citation closeness, which is the wrong signal for this task.
What is new here: a public test, written down in advance and run on topics from every domain of science, of whether such connections can be predicted before they happen; an open index of field-free problem descriptions, built by volunteers with whatever model they choose; and expert ratings that anyone can inspect. No step depends on one AI model or provider.
- Rerun the backtest: research/README.md.
- Use the data in any language: one table per freeze, one row per pair of topics, in the backtest release, with every column explained in docs/DATA.md.
- Write a client in your own language: docs/WRITE-A-CLIENT.md.
- Everything else: CONTRIBUTING.md.
Code: MIT. Data published by the project: CC0, derived from OpenAlex, which is also CC0.