Measure how a research field drifts
Bring a set of papers. An LLM you choose extracts the contribution claims from each abstract, the claims are embedded and clustered, and you get an interactive map of what the field produces, year by year.
1. Build a corpus
Paste DOIs or arXiv ids, upload a CSV or PDFs, or search OpenAlex, arXiv and Google Scholar. We find an abstract for each paper.
2. Extract claims
An LLM, on your own API key, lists the contribution claims in each abstract, for example “Instruction tuning improves zero-shot generalisation”.
3. Cluster and chart
SPECTER2 embeds the claims, UMAP and HDBSCAN group them, and each cluster gets a share per year: that change is the drift.
4. Explore and share
Open the five-view inspector (map, trends, clusters, compare, methods), download CSVs, and share a link or invite colleagues to comment.
Built on hamyrappy’s Drift Inspector
The method and the inspector are the work of hamyrappy, described in Drift Inspector: Exploring and Measuring Scientific Drift with Atomic Contribution Claims. This site runs that published code on your corpus. Like the paper, it works on abstracts by default.
You need an LLM API key (OpenRouter or any OpenAI-compatible endpoint).