What the site says

Users participate in blind side-by-side chat battles to evaluate and vote on responses from various large language models. The platform compiles these human evaluations into a crowdsourced Elo-based leaderboard for text, code, and vision capabilities.