Perplexity AI Releases WANDR: An Open Benchmark Evaluating Research Agents That Must Search Wide And Deep

What it does

Perplexity AI has launched WANDR, an open evaluation benchmark designed to test research agents’ ability to perform deep and wide search tasks. WANDR includes 500 complex, evidence-heavy challenges requiring agents to discover multiple qualifying entities and support each with clearly cited, verifiable evidence. This benchmark pushes AI systems beyond simple answer retrieval to emulate rigorous research with traceable sources.

Why it matters

For builders and operators working on AI search and research tools, WANDR sets a new standard for what these agents must achieve to be reliable. It moves the evaluation from just producing an answer to demonstrating evidence-backed findings, which addresses growing concerns about AI hallucinations and unverifiable results. Showing that a system can collect many relevant entities and cite trustworthy evidence raises the bar for real-world applications, especially in research, legal work, and due diligence.

Perplexity Search as Code currently leads on WANDR with a soft F1 score of 0.363 and a hard F1 of 0.133, signaling that there is still significant room for improvement. This performance illustrates how challenging it is for agents to cover the scope required for wide and deep research while maintaining accurate, citable sources.

Who it is for

WANDR primarily targets developers and teams building AI research agents, knowledge discovery tools, and data-driven decision support systems. It offers a concrete way to benchmark and improve agents tasked with information aggregation that demands transparency and factual backing. Investors and product leaders can use WANDR results to assess technical progress in search agents’ trustworthiness and real research use cases.

The catch

WANDR’s challenge reflects tough real-world conditions, and no current agent solves it comprehensively. The low hard F1 scores indicate difficulty in precisely matching all qualifying entities under strict evidence requirements. This shows that deploying large language model-based research tools in critical workflows still requires caution and rigorous testing against benchmarks like WANDR to avoid overconfidence in AI-generated evidence.

What to watch next

Advances with WANDR will push research agents toward better accuracy, broader coverage, and transparent sourcing. Watch for improvements from Perplexity and competitors seeking to raise their benchmark scores, which will signal maturing research AI technology. Beyond scores, integration of WANDR-style evaluation in operational settings will be key to validating whether agents truly reduce manual research time without sacrificing reliability.

AI Quick Briefs Editorial Desk