Work

Illustrated NeurIPS: A Navigable Map of 30,225 NeurIPS Papers (2020–2026) with Linked arXiv Records, Citations and Landmark Papers

Year: 2026
Type: preprint
Abstract: NeurIPS accepted 1,898 papers in 2020 and 9,111 in 2026. At that volume, reading the proceedings is no longer how researchers learn the shape of the field. We present Illustrated NeurIPS, an open dataset and browser-based map of all 30,225 papers accepted to NeurIPS from 2020 to 2026. Each paper is embedded from its title and abstract, placed on a two-dimensional map, and assigned to one of 285 fine topics nested in 30 areas, each with a short name and one-sentence description. We link 21,946 papers (72.6%) to their arXiv records, attach citation counts to 26,172, and add 1,035 highly cited "landmark" papers from other venues to the topics they belong to. Against NeurIPS's own subject-area labels, our two-stage clustering (k-means on UMAP features) agrees better than k-means on the raw embeddings (adjusted Rand index 0.086 vs. 0.065; paired-bootstrap 95% interval of the gain [0.018, 0.023]) and much better than TF-IDF clustering (0.056). Title-only search over the deployed embeddings retrieves the correct paper first 83.2% of the time among 26,632 abstracts (MRR 0.876, vs. 0.761 for TF-IDF). The map shows, for example, that LLM agent benchmarks and RL with verifiable rewards went from no NeurIPS papers in 2020–21 to 1.5% and 1.1% of 2025–26 papers, while generative adversarial networks fell from 1.0% to 0.01%. The site, data and code are public. Interactive map: https://vincentychu.com/neurips/. Code and evaluation: github.com/vinceychu/illustrated-neurips (scripts/neurips/, paper/).
Author: Chu, Vincent
Language: English
Open Access status: green
OpenAlex ID: W7235430579
Cites:
Cited by:
Related to: 10