TransMamba:A language model combining Transformer and Mamba

In recent years, State Space Models (SSMs) have achieved significant advancements in the field oflanguage modeling. With the advent of Mamba, these models have garnered even greater attention,surpassing Transformers in certain aspects. Despite Mamba’s unique advantages, Transformers remainindispensable due to their complex computational capabilities and proven effectiveness. This paperproposes a novel model that effectively combines the strengths of both Transformers and Mamba.Specifically, our model employs the Transformer’s encoder for encoding and utilizes Mamba as thedecoder for decoding. We introduce a feature fusion technique that integrates the features generated bythe encoder with the hidden states produced by the decoder. This approach effectively amalgamatesthe advantages of both Transformer and Mamba, resulting in enhanced performance. Extensiveexperiments on various language tasks demonstrate that our proposed model achieves competitiveresults, consistently outperforming existing benchmarks.