An open commons for machine learning research
More than three-quarters of a million papers distilled into 15 million extracted ideas, clustered into more than 25,000 research directions. Open without an account, and benchmarked in a public preprint.
lacuna
noun, Latin — gap
Lacuna is a living library of machine learning research, open to anyone and benchmarked in a public preprint (opens in new tab) presented at the ICML Workshop on AI for Research. It exists to tend and grow the open research culture that made modern AI possible, and to point at the gaps in our collective understanding.
What the map holds
Lacuna reads the papers it indexes and builds four layers on top:
| Layer | Count (preprint) | What it is |
|---|---|---|
| Paper pages | 733,795 | Markdown expositions of each paper's core ideas, some with the key figures extracted and kept |
| Concept elements | 15,259,720 | One- or two-sentence statements of a core idea, method, limitation, or observation, each linked to its papers |
| Research directions | 27,017 | Two-page syntheses of concept clusters; a recurring problem, method family, or opportunity area |
| Research proposals | 38,000 | Generated hypotheses, sampled from directions and their supporting papers |
The counts are the preprint's snapshot; the live site is already past 750,000 papers and 38,000 research directions, and keeps growing. Every generated item keeps links to the primary sources that support it, so wherever you land, the trail back to the original is one click.
Where the map comes from
The pipeline harvests and reconciles scholarly records from five sources, and it establishes stable identifiers before generating anything, so trails cannot dangle.
| Source | Role in the pipeline |
|---|---|
| OpenReview | The identity anchor for author pages: author-maintained records |
| arXiv, venue open-access pages | Paper harvest |
| OpenAlex, DBLP | Recall sources, not authorities |
The anchor choice is deliberate. External author IDs conflate researchers who share a name, a familiar problem for anyone who has found a stranger's papers on their own profile elsewhere; author-maintained OpenReview records do not.
Where the proposals come from: alien sampling
The map's research proposals are not free-form LLM ideation. They are produced with a published method we call alien sampling (arXiv:2603.01092 (opens in new tab)). In the paper, 16,068 peer-reviewed LLM papers are decomposed into a vocabulary of recurring "idea atoms," and two models score candidate combinations: coherence (would this hold together as research?) and availability (would any existing author community be positioned to produce it?). The sampler keeps combinations that score high on the first and low on the second, plausible under the structure of existing knowledge but unlikely under the distribution of existing researchers. The paper evaluates the sampler against frontier-model ideation runs, Claude Opus 4.7 and Gemini 3.1 Pro, each given the same atom vocabulary:
| What the paper measured | Alien sampling | Ideation baselines |
|---|---|---|
| Effective atom coverage | 72.5% | 20.6% and 10.5% |
| Effective atom vocabulary | 3.5–7× broader | — |
| Blind pairwise judging, average rank percentile | 0.633 | Claude 0.574, Gemini 0.495, random 0.295 |
| Recovering the atom sets of the following year's papers | 31× random sampling | — |
On the map, proposals are generated from research directions and their supporting papers, each one kept, in the Lacuna paper's words, as hypotheses for researcher review, with its trail back intact.
Benchmarked in public
The map is a published artifact, and so is the deep-research engine that runs over it. Our preprint (arXiv:2606.26246 (opens in new tab)) benchmarks both against research systems run head-to-head:
| Task (benchmark) | Metric | Lacuna | Baseline |
|---|---|---|---|
| Literature retrieval (LitSearch) | Recall@10 | 0.538 | OpenScholar v3: 0.424 |
| Survey reports (ReportBench-ML, 25 tasks) | Expert-reference hits | 99 | GPT-Researcher: 72 |
| Survey reports (ReportBench-ML) | Report quality (RACE) | 7.82 / 10 | GPT-Researcher: 5.24 / 10 |
| Related-work synthesis (Multi-XScience-CS/ML) | Score / 5 | 4.167 | OpenScholar: 3.467 |
| Literature QA (ScholarQA-CS-ML) | Rubric avg. | 0.694 | OpenScholar: 0.672 |
The baselines are the paper's: research systems, STORM among them. In the last row the ScholarQA margin is thin, and the paper reports it anyway. Full tables are in the preprint; review it the way you would review anything else.
Ask from any page
Lacuna is also the ground Althea works over. Open any paper, author, or direction and ask from right there; the page's context comes attached, because Lacuna already knows where you are. "What is the core claim of this paper?", "Find the weakest assumption in this method", and "Which neighboring papers should I read first?" all start from the page you are on, with its metadata, extracted concepts, figures, and neighboring papers carried over, the useful bits rather than the whole page dragged into the chat.
See it yourself
Each paper page carries the exposition, a link to the original on arXiv or the venue, figure captions that cite their exact source location, and the raw markdown one click away. Search runs on rank fusion across the map's layers, so a query surfaces directions and concepts as well as title matches. And the front page keeps one small tradition: Today in ML History, the field's anniversaries, dated and sourced. Worth the visit on Shannon's birthday.
Open to anyone without an account, human-readable and machine-readable. The commons belongs to all of us. Come help us tend it.
Open Lacuna Ask Althea