This paper addresses the challenge of long-tailed recognition, where the visual world exhibits a long-tailed distribution with a significant imbalance in the number of instances across classes. Traditional approaches often jointly learn representations and classifiers, but this paper decouples these two components to explore their individual contributions. The authors find that instance-balanced sampling is effective for learning high-quality representations, and that adjusting the classifier's decision boundaries can significantly improve long-tailed recognition performance. Extensive experiments on datasets like ImageNet-LT, Places-LT, and iNaturalist show that decoupling representation and classification can achieve state-of-the-art performance without the need for complex sampling strategies or additional modules. The proposed method outperforms both joint learning and various advanced techniques, demonstrating the effectiveness of decoupling in long-tailed recognition tasks.This paper addresses the challenge of long-tailed recognition, where the visual world exhibits a long-tailed distribution with a significant imbalance in the number of instances across classes. Traditional approaches often jointly learn representations and classifiers, but this paper decouples these two components to explore their individual contributions. The authors find that instance-balanced sampling is effective for learning high-quality representations, and that adjusting the classifier's decision boundaries can significantly improve long-tailed recognition performance. Extensive experiments on datasets like ImageNet-LT, Places-LT, and iNaturalist show that decoupling representation and classification can achieve state-of-the-art performance without the need for complex sampling strategies or additional modules. The proposed method outperforms both joint learning and various advanced techniques, demonstrating the effectiveness of decoupling in long-tailed recognition tasks.