**[AI Research Agent Capability Test: Perplexity Ranks First, But Only 1 in 7 Objects Fully Delivered]** According to monitoring by 动察 Beating, Perplexity has released the WANDR benchmark, specifically designed to test whether research agents can complete large-scale data collection tasks. The task requires first identifying a full set of objects, then supplementing each object with specified information and reliable sources. A total of 500 tasks involved over 170,000 records. Among the six systems tested, Perplexity Search as Code ranked first, but only 13.3% of the objects met the criteria of complete information and sufficient evidence. This translates to roughly 1 in 7 objects being fully delivered. The results indicate that while current research agents can locate a significant number of web pages, they still struggle with handling long lists and multi-field tasks, often missing objects, information, or evidence. [Original Link]