| title | Wikipedia Speedrun |
|---|---|
| emoji | 🏃 |
| colorFrom | green |
| colorTo | blue |
| sdk | docker |
| pinned | false |
| license | mit |
A benchmark comparing AI agents playing the Wikipedia Speedrun game - navigating from one Wikipedia article to another using only the links on each page.
Wikipedia Speedrun (also known as "Wiki Game" or "Wikipedia Golf") is a game where players start on one Wikipedia article and must navigate to a target article by clicking only the links within each page. The goal is to reach the target in as few clicks as possible.
Example challenges:
- Easy: Cat -> Dog (semantically close, many shared links)
- Medium: Coffee -> Minecraft (requires creative intermediate hops)
- Hard: Lamprey -> Costco Hot Dog (obscure start, specific target)
Run the Flask app and play in your browser:
python app.pyThen open http://localhost:7860 to play or watch AI agents compete.
We tested 15 AI agents across 35 unique problems, totaling 420+ games.
Points on the yellow Pareto frontier represent optimal cost-performance tradeoffs. Embedding models (green) are free while LLM models (blue) cost per API call.
Key Finding: Embedding-based agents achieve 94% win rates at zero cost, while LLMs reach 97% for minimal cost. The Pareto-optimal choices are bge-large (free, 94%) and gemini-2.0-flash ($0.001/game, 97%).
