Lacuna: A Research Map for Machine Learning, at ICML

We’re headed to Seoul to present Lacuna: A Research Map for Machine Learning at the ICML AI4Research workshop.
Lacuna is an auto-updating knowledge base for the machine learning community with author profiles, paper summaries with key figures, literature reviews of research directions, and proposals for new research. We built on the data structure produced for our Alien Sampling paper (in submission), and discovered that it was not only really good for producing diverse + high quality research ideas, but also really good for literature QA.
In Lacuna we extract key concepts from papers, cluster these into research directions, and produce markdown documents from the important papers in each cluster. We also (of course) do a little alien sampling and produce research proposals, too, then serve it all as web pages, raw markdown (for the agents), and MCP.
The most surprising part of this project was that we built Lacuna first as a means to sample more new ideas with Alien Sampling, and second because as ML researchers we wanted a good ML knowledge base. Anirudh Goyal pointed out (at ICLR this year) that it would probably be really good at literature search / QA as well. He was right (as usual). It took under a day to get the initial results surpassing ScholarQA's results on their AI-related questions, from AI2's Nature paper this year. Miles Li helped extending the results to the deep research setting and a couple other benchmarks. Pretty astonishing results, and we have a few more in store for the ICLR submission this fall!
- LitSearch Recall@10: 0.538 vs. 0.424 for OpenScholar v3
- Multi-XScience-CS/ML: 4.167 / 5 vs. 3.467 for OpenScholar
- ReportBench-ML: 99 expert-reference hits and 7.82 / 10 RACE report quality for Lacuna Deep Research, vs. 72 hits and 5.24 / 10 for GPT-Researcher
Congrats to all who've played a role in this so far Martin Weiss, Miles li, Alejandro H., Yacine Mkhinini, Chris Pal, Hugo Larochelle, and Nasim Rahaman.
Paper: arxiv.org/abs/2606.26246
Lacuna: lacuna.tiptreesystems.com
First posted on LinkedIn, July 1, 2026.


