RGPT is an adaptive boosting framework designed to enhance text classification performance by leveraging large language models (LLMs). The framework constructs a pool of strong base learners through iterative fine-tuning and sample distribution adjustment, then ensembles these learners to create a specialized text classification LLM. RGPT outperforms 8 state-of-the-art pre-trained language models (PLMs) and 7 state-of-the-art LLMs on four benchmark datasets, achieving an average improvement of 1.36%. It also surpasses average human classification performance. The framework's effectiveness is demonstrated through comprehensive experiments, showing that RGPT's performance continues to improve with more iterations. RGPT's approach is less sensitive to prompts and more stable across tasks, making it a robust solution for text classification. The study highlights that RGPT can universally boost various base model structures, pushing the limits of LLM capacity for text classification. Key contributions include exploring the ongoing research value of text classification in the era of LLMs, proposing RGPT as an adaptive boosting framework, and demonstrating its effectiveness in zero-shot text classification across four datasets. The framework's recurrent ensembling approach leverages historical outputs to improve classification accuracy and generalization. RGPT's performance is validated through human evaluation, showing it outperforms human annotators in accuracy and efficiency. The study also addresses overfitting risks and demonstrates the model's ability to generalize to new instances. Overall, RGPT represents a significant advancement in text classification by effectively utilizing LLMs and enhancing their classification capabilities through adaptive boosting.RGPT is an adaptive boosting framework designed to enhance text classification performance by leveraging large language models (LLMs). The framework constructs a pool of strong base learners through iterative fine-tuning and sample distribution adjustment, then ensembles these learners to create a specialized text classification LLM. RGPT outperforms 8 state-of-the-art pre-trained language models (PLMs) and 7 state-of-the-art LLMs on four benchmark datasets, achieving an average improvement of 1.36%. It also surpasses average human classification performance. The framework's effectiveness is demonstrated through comprehensive experiments, showing that RGPT's performance continues to improve with more iterations. RGPT's approach is less sensitive to prompts and more stable across tasks, making it a robust solution for text classification. The study highlights that RGPT can universally boost various base model structures, pushing the limits of LLM capacity for text classification. Key contributions include exploring the ongoing research value of text classification in the era of LLMs, proposing RGPT as an adaptive boosting framework, and demonstrating its effectiveness in zero-shot text classification across four datasets. The framework's recurrent ensembling approach leverages historical outputs to improve classification accuracy and generalization. RGPT's performance is validated through human evaluation, showing it outperforms human annotators in accuracy and efficiency. The study also addresses overfitting risks and demonstrates the model's ability to generalize to new instances. Overall, RGPT represents a significant advancement in text classification by effectively utilizing LLMs and enhancing their classification capabilities through adaptive boosting.