Researchers have introduced ConvergeFlow, a new type of continuous diffusion and flow-based language model. Unlike existing models, ConvergeFlow constrains the data predictor to valid token embeddings, allowing for direct token prediction without additional training. The team proves that this approach converges to valid token embeddings under certain conditions. Experiments show that ConvergeFlow achieves performance on par with other continuous and discrete diffusion LMs. The code is available online. --- Why it matters: This matters because it improves the efficiency of language models by eliminating the need for a separate decoder, which can be computationally expensive. It also opens up new possibilities for using flow-based paradigms in language modeling. Source: https://arxiv.org/abs/2608.23551