Diffusion models are usually built to generate from one fixed target distribution, or from a family of distributions tied to conditioning variables chosen during training. That setup makes it hard for a trained model to switch to a task it has not seen before.
The paper proposes transformer-based in-context score operator learning as an alternative. Rather than retraining for each new target, the model learns to infer the score operator from context supplied at inference time, so it can adapt to previously unseen distributions on the fly.
The idea points toward more flexible, task-agnostic generative models. The released abstract stops before reporting experimental details, so the available source does not yet indicate how the approach performs in practice.