Perplexity open-sources internal research benchmark WANDR

On 07-15, Perplexity open-sourced WANDR, its internal benchmark and eval/RL environment for measuring agentic search and "deep and wide research" capabilities. What makes it notable is that this is not a demo benchmark for toy tasks, but an internal R&D tool synthesized from production traces to approximate the real usage distribution — now released to outside researchers and developers. It comprises 500 tasks.

Key details

Multiple posts describe WANDR as the benchmark and accompanying eval/RL environment Perplexity uses internally to build research capabilities. @andersonbcdefg relayed that the environment is synthesized from production traces to stay close to real distribution rather than toy tasks, and uses only "weak…" (text truncated). @denisyarats put the scale at 500 tasks. @AccBalanced added that the benchmark has been used for work around GLM 5.2 J-space reproduction.

Evaluation focus

The posts converge on what "wide research" means: not just surfacing a few relevant results, but finding all qualifying results and providing evidence for each. @dmitry140 elaborated that these tasks target web or internal data, requiring retrieval, summarization, and verification across hundreds to thousands of entities — closer to complex research workflows than ordinary QA.

Background and outside interpretation

Several reposts note WANDR was the internal evaluation tool used while building Perplexity's deep and wide research capabilities and related Perplexity Computer features. @heathercmiller argued that companies able to encode their value proposition into their evals and environments will have an edge — which helps explain why Perplexity chose to open-source the evaluation system rather than merely demo product capability. @inductionheads relayed Suhail's view that Perplexity is particularly good at this kind of benchmark and, after digging in, leads the field in this direction.

2026-07-15 ~ 2026-07-15 · 9 related posts

• Perplexity Open-Sources Internal Benchmark WANDR — camerontstow · 2026-07-15
• Perplexity Open-Sources WANDR Research Benchmark — inductionheads · 2026-07-15
• Perplexity Releases WANDR Research Benchmark — denisyarats · 2026-07-15
• Perplexity Open-Sources Agentic Search Eval Environment — andersonbcdefg · 2026-07-15
• Perplexity Open-Sources Internal Benchmark WANDR — heathercmiller · 2026-07-15

4 near-duplicate retellings: dmitry140 · inductionheads · AccBalanced · denisyarats