<p>The prevalence of long-tailed distributions results in a significant disparity in sample sizes, often causing models to favor well-represented head classes while struggling with the recognition of less common tail classes. This challenge is further compounded by the presence of unknown classes in real-world data, which exhibit high uncertainty and can overlap with tail classes in the feature space, complicating their identification. In response, we introduce an innovative approach to long-tailed classification tailored for open-set conditions, grounded in meta-uncertainty calibration. Our method employs a multi-expert framework, informed by extreme value theory, and augmented with evidence theory to mitigate the confusion between tail and unknown classes. Initially, extreme value theory is applied to refine the uncertainty estimates for class predictions, thereby bolstering confidence in the model’s predictions for challenging classes. Subsequently, evidence theory is leveraged to amalgamate the predictions from individual experts, enhancing the model’s overall discernment for complex classifications. Additionally, we have devised a strategy that dynamically adjusts the number of experts engaged in the decision-making process based on the difficulty of the classification task, thereby optimizing computational efficiency. Our experiments substantiate the efficacy of our method in recognizing both tail and unknown classes effectively.</p>