What it is
Benchmark for retrieving prior papers that genuinely inspired new research, using author-verified questions, positives, hard negatives, and rationales. 191k-paper corpus with 894 author-verified research questions from 207 CS projects, plus relevance labels, hard negatives, and rationale annotations for core-paper and subfield retrieval.
Key results
- 190,896 paper title+abstract documents
- 894 queries from 207 projects by 184 researchers
- CC-BY-NC-4.0; 207 core and 687 subfield query judgments
Useful for evaluating retrieval systems that find papers likely to inspire a new research project or idea thread.