In-Context Adaptation

Simple Definition

In-context adaptation is changing an AI system’s behavior using examples, instructions, demonstrations, or feedback inside the current context, without updating the model’s weights.

Plain-English Explanation

This is the umbrella idea behind a family of techniques you may already know by more specific names. When you give a language model a few examples of the format you want, that’s few-shot prompting. When you give it a clear instruction with no examples at all, that’s zero-shot prompting. Both are specific, text-based instances of in-context adaptation: the model’s behavior shifts based on what’s currently in front of it, not because it was retrained.

The same underlying idea shows up outside of text, too. A robot that watches a short physical demonstration and immediately attempts a similar action is adapting in context as well, just through a different modality. See Physical Prompting for that specific case.

Example

An LLM given three examples of a desired writing format follows that pattern for the rest of the current conversation. Separately, a robot watches a three-second demonstration of how to pick up an object and immediately attempts the task the same way. Both are adapting behavior using what’s currently available to them, not because anything about the underlying model changed.

An Important Limit

The adaptation is usually temporary. It disappears once the relevant context is gone, unless the lesson gets saved somewhere persistent, a memory system, a context file, or an actual training step. Close the conversation, and the model has no memory of the examples you gave it unless something explicitly carried that information forward.

In-Context Adaptation vs. Continual Learning

These sit on opposite ends of the same spectrum. In-context adaptation keeps the model’s weights unchanged, is temporary by nature, and depends entirely on what’s currently in context. Continual learning is the deeper version: the system actually learns from new experience over time, potentially updating model parameters or some other persistent learning mechanism, so the change outlives the current session.

Few-Shot Prompting, the specific technique of adapting behavior with a handful of examples
Zero-Shot Prompting, adapting behavior from instructions alone, with no examples
Continual Learning, the deeper, persistent version of learning from experience
Physical Prompting, in-context adaptation applied to robots learning from a physical demonstration
Fine-Tuning, permanently baking a behavior in, rather than adapting to it temporarily

For how this compares to genuine continual learning, see Continual Learning Explained.