WANDR Benchmark Evaluates AI Research Agents on Complex Tasks

A new benchmark called WANDR (Wide ANd Deep Research) has been introduced to rigorously assess the capabilities of AI research agents. This benchmark comprises 500 complex data-collection tasks that mirror real-world scenarios, demanding agents to both broadly discover entities and deeply investigate each one using coordinated web searches. The ultimate goal is to produce independently verifiable records, complete with supporting sources and excerpts. WANDR's tasks are structured as hierarchical qualification keys, specifying the number of entities, relationships, and evidence required at each level. This design supports diverse professional workflows such as market mapping, due diligence, and talent sourcing, with target record counts ranging from dozens to thousands. Unlike benchmarks with static answer sets, WANDR employs dynamic, task-specific judges that refetch cited pages to verify records against current facts. Initial evaluations of six production research systems on WANDR indicate that the benchmark is highly challenging. The strongest system achieved only 0.363 soft F1 and 0.133 hard F1, with performance significantly degrading as the volume and depth of the hierarchy increased. Key bottlenecks identified include incomplete discovery, missing enrichment, and inadequate evidence construction, suggesting ample room for improvement in current AI research agents.