Abstarct: In recent years, the amount of textual data, which is considered a source of information, has increased significantly. In order to make better use of this information in a reasonable time, summarization is necessary. Language models are one of the most powerful summarization tools available today. But it is not possible to find a model which can be claimed to produce the best summary for any text. Different language models look at the text from different aspects, and as a result, the summary produced by each one may be of high or low quality. One way to solve this problem is to use ensemble learning, which can happen in the model space or in the output space. Considering the high processing requirement in language models baxsed on transfer learning, it seems more optimal to use the latter. When the goal is to combine several texts (summaries here) and create an output from them, it is not easy to provide a solution to combine the outputs. For this purpose, a neural network was used to produce an ensemble model to combine the outputs of other networks. To this end, five pre-trained transformers were used to generate outputs, using these outputs a new dataset was created and used to train a transformer to improve the summaries. Considering that the output of each model captures different aspects of the text, its goal is to combine the strengths of each model by combining these outputs. The proposed model was evaluated with various evaluation criteria, performed better in all criteria and was found to produce significantly better output than any baxseline model.