Reproduce “The Leaderboard Illusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shivalika Singh, Yiyang Nan, Alex Wang, Daniel D'Souza, Sayash Kapoor, Ahmet Üstün, Sanmi Koyejo, Yuntian Deng and 5 more
The authors audit how Chatbot Arena’s private testing, battle exposure, and model removals can affect its rankings. They compare identical Aya-Vision-8B checkpoints that received scores of 1052 and 1069. In one 7B fine-tuning setup, Arena-data training raised ArenaHard win rate from 23.5% to 49.9%, but did not improve MMLU performance, limiting the result to an Arena-distribution test.
Key insights
- Undisclosed private testing with selective variant release systematically inflates model scores on leaderboards like Chatbot Arena, violating unbiased sampling assumptions.
- Proprietary model providers gain disproportionately more battle data compared to open-weight and open-source models due to unequal sampling rates and deprecation policies.
- Access to Chatbot Arena-specific data enables significant performance gains on the benchmark itself (e.g., doubling win-rates on ArenaHard), but these gains do not necessarily generalize to out-of-distribution benchmarks, indicating overfitting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.