Reconstruction of gene regulatory networks (GRNs) from large-scale expression data can mine the potential causality relationship among the genes and help understand the complex regulatory mechanisms. It is of utmost interest and has become a challenging computational problem for understanding the complex regulatory mechanisms in cellular systems. For the past decades, numerous theoretical and computational approaches have been introduced for inferring the GRNs. However, all existing methods of inferring GRNs from gene expression profiles have their strengths and weaknesses. In particular, many properties of GRNs, such as topology sparseness and non-linear dependence, are generally in regulation mechanism but are seldom taken into account simultaneously in one computational method. Some information theory algorithms do not recover the true positive edges that may have been deleted in an earlier computing process. These interaction relationships may reflect the actual relationship of genes. To overcome these disadvantages and to further enhance the precision and robustness of inferred GRNs, we presented an ensemble method, to infer GRNs from gene expression data by adopting two strategies of resampling and arithmetic mean fusion in this work. In this algorithm, the jackknife resampling procedure was first employed to form a series of sub-datasets of gene expression data, then the conditional mutual information was used to generate the corresponding sub-networks from the sub-datasets, and the final GRN was inferred by integrating these sub-networks with an arithmetic mean fusion strategy. Compared with those of the state-of-the-art algorithm on the benchmark synthetic GRNs datasets from the DREAM3 challenge and a real SOS DNA repair network, the results show that our method outperforms significantly LP, LASSO and ARANCE methods, and has a high and robust performance.