How Lacuna is made
Lacuna is a research map of machine learning built by Tiptree Systems. It turns scholarly records and paper content into linked, inspectable objects: paper summaries, concept elements, research directions and research proposals. Each generated item keeps links to the records and papers that produced it. Today the map covers 2,250,288 papers, 215,519 authors and 28,588 research directions, and is updated daily. The construction is described in full in the Lacuna paper; this page is the short form.
1. Sources
Lacuna harvests and reconciles records from OpenReview, OpenAlex, DBLP, arXiv and the open-access pages of the major machine learning venues. Paper files or URL replicas are attached to stable paper records, so the identity of a paper is separate from where its file can be fetched. New arXiv papers by the researchers Lacuna follows are added every day. Bibliographic facts on a page (title, authors, venue, year, identifiers) come from these sources and are not generated.
2. Records and identities
Author identities are anchored on OpenReview, whose author-maintained profiles give a cleaner signal than external identifiers alone, which can conflate researchers who share a name. For each OpenReview-profiled author, the pipeline harvests that author's OpenReview papers, canonicalizes aliases and builds a co-author neighbourhood from OpenReview papers only. DBLP and OpenAlex are used as recall sources, not as authorities: a paper found through an external identifier is kept for an author only when it reconciles to an existing OpenReview paper or its author list overlaps that trusted neighbourhood. Venue pages and venue listings aggregate the same records by conference and year.
3. Generated layers
Four layers are generated by large language models from the records above. Every one of them links back to its sources.
| Layer | What it is | Made from |
|---|---|---|
| Paper summary | A markdown summary of a paper's main technical ideas; a second form adds figures extracted from the PDF and selected for context. | Paper text and metadata; figures detected with RT-DETR. |
| Concept element | A one- or two-sentence statement of a method, observation, limitation or empirical finding, with the snippet it came from. | Paper summaries. |
| Research direction | An essay of about two pages on a recurring problem, method family or opportunity area, with its papers, authors, venues and neighbouring directions. | Concept elements embedded and clustered with HDBSCAN; the cluster's summaries and concepts. |
| Research proposal | A candidate hypothesis sampled from research directions and their supporting papers. | Directions, paper summaries with figures, concept elements. |
Author overviews on author pages are generated the same way, from the author's papers and research directions, and say when they were last updated.
4. Research directions
A research direction is a cluster of concept elements: recurring technical problems, methods and opportunity areas found across many papers. Its page carries the generated essay, the papers and authors that contribute to it, its venues, related proposals and neighbouring directions. Direction pages are versioned. The byline under the title gives the version number, the date it was first published and the date it was last updated; earlier versions stay at their own addresses so a citation keeps pointing at the text that was cited.
5. Research proposals
Proposals are generated hypotheses: each one names the directions and papers it draws on, so a reader can follow the evidence from the proposal back to the underlying paper pages. They are offered as candidates for investigation, not as findings.
6. Provenance, versions and addresses
- Provenance. Each generated item keeps links to the papers, metadata records, concept elements and directions that produced it.
- Sources. Every paper page links to the paper on arXiv or at its venue.
- Stable addresses. Swap
arxiv.orgforlacuna.tiptreesystems.comin any arXiv address to reach a paper's page. A retired entry returns 410, never a different page. - Machine-readable twins. Each content page has a Markdown version at
the same address with
/mdappended, and the Lacuna MCP server exposes the map to agents.
7. Evaluation
The paper evaluates the map on literature search and grounded question answering. Across LitSearch, Multi-XScience-CS/ML and ScholarQA-CS-ML, Lacuna outperforms OpenScholar, with the largest gain on LitSearch retrieval (Recall@10 of 0.538 against 0.424 for OpenScholar v3). On 25 ReportBench-ML survey tasks, Lacuna Deep Research, a multi-stage report agent over the map, reaches a citation F1 of 0.052, citation precision of 0.339, 99 expert-reference hits and a RACE report quality of 7.82 out of 10, against 0.039, 0.290, 72 and 5.24 for GPT-Researcher. Details and the per-claim audit are in the paper.
8. What can be wrong, and corrections
Generated text can misstate a result, miss a paper that belongs in a direction, or attach a paper to the wrong author when sources disagree. That is why every claim on the site sits next to a link to its source: check the paper, not the summary. Author and direction pages have a Report a mistake link at the bottom, and a researcher can claim their page to correct it directly. For anything else, use the feedback form; we reply to every report.
9. People
Lacuna is built by Tiptree Systems in Montréal. The Lacuna paper is by:
- Martin Weiss, Tiptree Systems, Mila and Polytechnique Montréal
- Miles Q. Li, Tiptree Systems and McGill University
- Alejandro H. Artiles, Tiptree Systems
- Yacine Mkhinini, Tiptree Systems
- Chris Pal, Mila and Polytechnique Montréal
- Hugo Larochelle, Mila and Université de Montréal
- Nasim Rahaman, Tiptree Systems
Citing Lacuna
Weiss, Li, Artiles, Mkhinini, Pal, Larochelle and Rahaman, "Lacuna: A Research Map for Machine Learning", arXiv:2606.26246 (2026).
To cite a page, link to its address; direction pages carry a version in their byline.