Researchers tested language models on the Wikipedia Game, a task requiring information retrieval from large document collections. They compared sentence embedding models and LLMs across languages and datasets, providing code for building and navigating local graphs of Wikipedia pages and paper citations.