Indie Dev Builds LLM Consensus Leaderboard Using Esports Ranking Algorithms

数字生命卡兹克 · wechat · 2026-08-10

The author spent a weekend developing AIHOT, an aggregated LLM leaderboard that compiles data from various major evaluation boards to calculate a unified "consensus score."

The main challenge was fairly merging different leaderboards. Simple averaging fails to account for varying numbers of models, score margins, and missing data. To solve this, the author stepped outside traditional approaches and used AI to find inspiration from educational grading and esports matchmaking.

Ultimately, the system adopted the Bradley-Terry paired comparison algorithm, similar to the Elo system in esports. It converts direct model comparisons across boards into win/loss relationships, using common opponents to build a global network that infers latent model abilities. The system also incorporates priors and anchor models to curb small-sample variance and normalize scores to a 100-point scale. The author transparently published the algorithm details, acknowledging it as an objective aggregation effort rather than an absolute standard.

Original post →

More from Apps

Apps channel →