AI for Science workshops 2021–2025

Small corpus: the method was validated at about 750 papers per year; clusters and drift on fewer papers are noisy.

Method and citation

Method

This analysis runs the original Drift Inspector methodology: Atomic Contribution Claims extracted from titles and abstracts by an LLM, SPECTER2 embeddings, UMAP and HDBSCAN clustering, and per-year document frequency.

Vsevolod Karimov, Stepan Ostarkov, Anastasia Poroshina, Anatoly Frolov, Alexander Panchenko. Drift Inspector: Exploring and Measuring Scientific Drift with Atomic Contribution Claims. 2026. arXiv:2609.39710 · code (MIT)