Perplexity open sources WANDR benchmark for testing research agents
Releasing an internal tool to measure how well AI agents search and reason
@perplexity_aiTL;DR:
- Perplexity is open-sourcing WANDR, a benchmark they used in-house
- It checks if agents can search widely and dig deep on complex tasks
- Existing benchmarks fall short on this mix of breadth and depth
Headline
Perplexity is open-sourcing WANDR, their internal benchmark for testing research agents that go both deep and wide.
Summary
Perplexity built WANDR to improve research features in their own AI. The benchmark pushes agents to pull information from many sources while also working through tricky, multi-step questions. As more tools claim to do agent-style research, having a standard way to judge quality beyond just getting the answer right starts to matter.
Analysis
Companies are putting more effort into agent benchmarks lately, especially ones that mix web search, tool use, pulling sources together, and step-by-step reasoning. By releasing this one, Perplexity looks like it's helping set standards instead of just selling another search product. It gives developers a way to compare agents on tasks that need both broad exploration and real analysis, something most current LLM tests don't handle well. Open-sourcing eval tools has become common among AI labs that want to steer how the field measures agent workflows.
Impact Assessment
Significance: Medium Categories: Open Source, AI Research, Developer Tools