Contrary to what the leaderboard lists as the human score, their technical paper implies a human baseline of ~48%.