Language models can learn new tasks from a handful of demonstrations, without any weight updates. That is the appeal of in-context learning. But the mechanism is expensive: every inference has to process the entire set of examples, so deployment becomes inefficient.
A new arXiv preprint, "Capturing In-Context Learning Dynamics with Task Operators," proposes task operators as a way to capture how in-context learning unfolds. The abstract frames task operators as a response to the cost problem, though it does not spell out the full formulation.
Because there is only one source, there are no conflicting findings to note. The paper's contribution, as described, is to offer a representation of in-context learning dynamics in service of that efficiency goal.