Daily AI Paper

Today’s standout paper is ConvergeFlow, a new way to build language models using continuous flows instead of the usual next-token classifier. The problem it tackles is a subtle one: continuous diffusion and flow models can generate smooth trajectories, but those trajectories do not always land exactly on valid token embeddings, so systems often still need a cross-entropy-trained decoder to turn them into text. ConvergeFlow changes that by constraining the model’s predictions to the convex hull of token embeddings and training only with a flow-matching objective. The authors prove that, under reasonable conditions, the flow converges to real token embeddings even when the predictor is imperfect. That matters because it removes a major engineering crutch and makes direct token prediction possible in a cleaner generative framework. In experiments on OpenWebText, the model is competitive with both continuous and discrete language models, suggesting that flow-based language generation may be more than a theoretical curiosity.