Benchmark
WANDR
Can a web research agent find a large set of qualifying records and support each claim with evidence?
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
Perplexity’s 500 tasks require agents to identify qualifying entities and document facts about them. A task can require evidence for both an entity and each of its related records.
Methodology
Task-specific judges revisit cited pages. Soft metrics credit partial coverage, while hard metrics require all requested evidence for a member. Web sources can change between runs. The paper reports a separate 45-task subset for comparisons across effort settings.