This paper investigates transfer learning for nonparametric regression, focusing on the non-asymptotic minimax risk and developing a novel estimator called the confidence thresholding estimator. The study reveals two unique phenomena in transfer learning: auto-smoothing and super-acceleration. The confidence thresholding estimator achieves the minimax optimal risk up to a logarithmic factor. A data-driven algorithm is proposed that adaptively achieves the minimax risk across a wide range of parameter spaces. Simulation studies and a real-world example demonstrate the effectiveness of the proposed method. The paper also extends the results to multiple source distributions and discusses the theoretical and practical implications of transfer learning in nonparametric regression. The results show that transfer learning can significantly improve estimation accuracy, especially when the bias strength is small and the source domain has a smoother mean function or larger sample size. The adaptive confidence thresholding algorithm is shown to achieve the minimax optimal risk up to a logarithmic factor, demonstrating its effectiveness in various scenarios.This paper investigates transfer learning for nonparametric regression, focusing on the non-asymptotic minimax risk and developing a novel estimator called the confidence thresholding estimator. The study reveals two unique phenomena in transfer learning: auto-smoothing and super-acceleration. The confidence thresholding estimator achieves the minimax optimal risk up to a logarithmic factor. A data-driven algorithm is proposed that adaptively achieves the minimax risk across a wide range of parameter spaces. Simulation studies and a real-world example demonstrate the effectiveness of the proposed method. The paper also extends the results to multiple source distributions and discusses the theoretical and practical implications of transfer learning in nonparametric regression. The results show that transfer learning can significantly improve estimation accuracy, especially when the bias strength is small and the source domain has a smoother mean function or larger sample size. The adaptive confidence thresholding algorithm is shown to achieve the minimax optimal risk up to a logarithmic factor, demonstrating its effectiveness in various scenarios.