Research teams today delegate competitive mapping, due diligence and literature reviews to AI agents, but most benchmarks only check a single correct answer. They do not measure whether an agent can gather a large, evidence‑backed collection of facts—a real‑world requirement for knowledge work. Perplexity’s new open benchmark WANDR (Wide ANd Deep Research) fills this gap by testing both breadth and depth in a single evaluation.
WANDR presents 500 realistic data‑collection tasks derived from actual production patterns, not synthetic prompts. Each task uses a qualification key hierarchy that specifies how many top‑level entities must be found, how many sub‑entities per entity, and how many supporting sources per sub‑entity. For example, a task may require 70 US‑based companies, each with a CEO or CFO appointment announced between March and April 2026, plus one authoritative appointment page and a listing‑authority page per company. This creates a clear, measurable target: 140 source‑backed records that must be discovered, enriched and verified.
Grading in WANDR is claim‑centric. Every submitted record includes an item, URL, excerpts and an answer. During evaluation the grader re‑fetches the page, checks that it is usable and in scope, and confirms that the excerpts truly support the claim. Binary verdicts for each record roll up through the hierarchy to produce precision, recall, soft F1 (partial credit) and hard F1 (full subtree correct). The benchmark shows that even the best systems achieve only modest scores—soft F1 around 0.36 and hard F1 around 0.13—revealing major gaps in discovery, evidence verification and handling of deeper hierarchies.
The practical takeaway for teams building research agents is to treat wide discovery and deep evidence collection as separate, measurable stages. Focus first on improving recall at the top level (finding enough qualifying entities), then strengthen the deep verification step (ensuring each entity is backed by reliable, correctly cited sources). Using a programmable agent architecture that can express retrieval, filtering, fan‑out, joins and stopping logic helps manage the repeated operations needed for large collections without overloading model context.
By adopting WANDR‑style evaluation, developers can pinpoint where their agents lose precision or recall, iterate on targeted improvements, and move closer to systems that reliably produce comprehensive, evidence‑rich research outputs. #AI #Product #Research #Agents #Benchmark #LLM