Can Language Models Play the Wikipedia Game?
The tested LLMs could match human navigation in the Wikipedia Game, but did not significantly beat it.
The paper uses link-only navigation through Wikipedia as a proxy for retrieval across large document collections. It compares embedding models and LLMs across five Wikipedia languages and five academic-literature datasets, using local Wikipedia and OpenAlex citation graphs. GPT-o3 mini, GPT-4o, QwQ:32B, and DeepSeek-R1:70B were among the strongest systems. Results shifted by language and topic, which the authors flag as relevant for practitioners using these methods. ArXiv · AI/CL/LG's note
The paper uses link-only navigation through Wikipedia as a proxy for retrieval across large document collections. It compares embedding models and LLMs across five Wikipedia languages and five academic-literature datasets, using local Wikipedia and OpenAlex citation graphs. GPT-o3 mini, GPT-4o, QwQ:32B, and DeepSeek-R1:70B were among the strongest systems. Results shifted by language and topic, which the authors flag as relevant for practitioners using these methods. ArXiv · AI/CL/LG's note
score 4