in-context skill acquisition

In-context skill acquisition is the striking ability to pick up a brand-new task from the prompt alone, with no retraining. Define a made-up tagging scheme, show two or three examples, and the model applies it to a fresh case. Invent a small artificial language with a few translation pairs and it will translate a new sentence. The skill lives only in the current conversation: nothing about the model changed, yet for the length of that prompt it behaves as if it had been taught.

What is really happening is that the examples in the prompt steer the model toward a pattern it can already represent, then it continues that pattern. This is why it feels like learning but is better described as rapid adaptation: the moment the context is gone, so is the skill, and a follow-up question in a fresh chat finds the model knowing nothing of your scheme. It also explains the limits. The prompt can only summon abilities latent in what the model already learned; it cannot install a capability that pretraining never built.