Evaluation & BenchmarksShivalika Singh et al. (Cohere Labs and others)Cohere

Cohere Labs and others publish The Leaderboard Illusion

Published The Leaderboard Illusion — Google receives 19.2% and OpenAI 20.4% of Arena data, while 83 open-weight models combined receive 29.7%. Providers can test multiple private variants and then retract scores

Why it matters

Demonstrated the structure of leaderboard overfitting. The reason Arena rankings should not be read as a direct measure of capability

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2504.20879

Added in Week 35, 2026 · See that week

Same week

Qwen3OfficialMem0 publishes its paperPaper

Nearby on the same axis

Controversy over unreleased experimental Llama 4 Maverick behind Arena scorePressARC Prize announces ARC-AGI-2 and ARC Prize 2025OfficialEpoch AI posts a clarification on how FrontierMath was made and is usedOfficialTechCrunch reports criticism of Epoch AI for not disclosing OpenAI fundingPress