In-Context Learning
A model learning a task temporarily from instructions and examples in the prompt without updating its weights.
In plain language
In-context learning is one of the most practical LLM abilities: the same model can classify, extract, rewrite and code through different prompts. It relies on attention mapping examples to the input rather than changing knowledge. It is sensitive to example order, format and wording, and can learn spurious correlations. Production tasks still need a fixed evaluation set; one good demo is not evidence.
Example
Give the model three complaint-to-response examples and it can respond in a similar style to a new complaint without any weight update.
Recommended tools
Related terms
Few-Shot
Providing a small number of input-output examples in the prompt so the model can imitate the pattern.
Prompt Engineering
Designing instructions, examples and constraints so a model reliably produces the desired output.
Zero-Shot
A model performing a task without any examples of that task in the prompt.
Agent Memory
A mechanism for an agent to store and retrieve user preferences, past tasks and key facts across sessions.