AllenAI researchers have introduced DiScoFormer, a transformer architecture that jointly handles two core tasks in generative modeling: estimating probability densities and computing score functions (gradients of log-density). Unlike conventional methods that train separate networks for each task, DiScoFormer learns a single representation that works for both, across a range of target distributions.\n\nThe key innovation is that the model is not tied to a specific distribution. Instead, it takes a distribution as part of its input, allowing it to adapt to new data distributions at inference time without retraining. This is a departure from typical score-based and normalizing-flow models, which are usually fitted to one dataset.\n\nThe authors report that DiScoFormer achieves competitive performance on standard benchmarks while being more flexible. By unifying density and score estimation, it could simplify the architecture of generative systems that currently rely on multiple components. The work is a step toward more general-purpose probabilistic models.