Navigating to a page by following links alone tests information retrieval. The best models match human performance, but none significantly outperform it.
Researchers are comparing language models and sentence representation models in the game Wikipedia: joining a target article without leaving the links of the encyclopedia. They extend this test to five languages and five collections of scientific literature, with citation graphs from OpenAlex.
System performance and ranking change by language and domain. The best models achieve success comparable to human participants, with no statistically significant benefit. The code makes it possible to reconstruct these journeys locally, without relying on APIs from third-party services.
Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.
Read the full article on the original source
Go to the source →