Even the Best Research Agent Only Gets 1 in 7 Objects Right

Perplexity just dropped a new benchmark called WANDR, designed to test whether research agents can handle large-scale data gathering. The task: find a list of objects, then fill in specific info and reliable sources for each one.

500 tasks across more than 170,000 records. Six systems were put through the wringer. Perplexity's own Search as Code came out on top — but even it managed to fully complete just 13.3% of objects with both information and evidence intact. That works out to roughly one in seven.

The results show that today's research agents can find plenty of web pages, but when it comes to long lists and multi-field tasks, they still regularly miss objects, drop details, or fail to cite sources.