According to CoinMeta, Arena has launched AutoEval, which utilizes a reward model to simulate user voting in order to quickly generate leaderboard scores for the new model. After the new model is released, it can obtain an estimated ranking in less than 1 hour, and the scores will be marked as AutoEval. Subsequently, these scores will be corrected based on the results of actual human voting. This system is trained using millions of sets of real user preference data, and during evaluations, it compares the quality of responses from different models, calculating scores according to the original leaderboard rules of Arena. Currently, it supports evaluations for text, visual, image generation, and code models. Backtesting with Arena has shown that the correlation between AutoEval and subsequent human rankings exceeds 0.98, and the accuracy rate in determining winners when there is a score difference of more than 10 points is over 90%.