Sentence fusion remains a challenging task for summarization systems. Multiple compatible sentences must be selected, then important content must be extracted from those sentences, and finally the content must be merged together into a natural language sentence. In this paper, we analyze the outputs of five state-of-the-art neural abstractive summarizers, focusing on summary sentences that are formed by sentence fusion. We find that system sentences are mostly grammatical, but often fail to remain faithful to the original article. Additionally, we propose a new framework for generating summary sentences that divides sub-tasks between two models: one which selects sentences and highlights words and phrases within those sentences, and the other generates a sentence using the previous model’s outputs as supplementary information. We achieve promising results using this distribution of tasks, which allows the generation model to better focus on fusing sentences effectively.