CrossSearch Benchmark
Explore score history ↗Evaluation-only benchmark: 100 professional tasks where an agent must reconcile files in a local workspace with public web sources, graded on atomic rubrics for full credit (reported as pass@1 and pass^3). The dataset preserves rollout traces, attachments, reference answers, rubric judgments and model rewards, and documents native and alternative search-engine conditions separately. A reward may be absent for a limit outcome, and results from different conditions, judge versions or task-set revisions are not interchangeable. The authors' blog reports a leaderboard and dates the release 2026-11-12, but the dataset card states no release date and this registry cites only the benchmark URL, so the entry carries no released field; the blog also reports two metrics, so no score rows are added.
- Evaluates
- professional
- Openness
- unknown
- Importer
- Model reports
CrossSearch links
No reported scores are on record for this benchmark yet.