What it is
Few shot in context learning is the ability of a large language model to pick up a task from a handful of examples placed inside the prompt, with no weight updates at all. The model infers the pattern and applies it to a new input.
How it works
You show the model the pattern you want and then a fresh case.
- Each example pairs an input with its desired output
- The new query is added at the end, and the model continues the pattern
- Zero shot gives no examples, one shot gives one, and few shot gives several
Nothing is trained here. The examples only steer the model at inference time, which is why it is called in context learning rather than fine tuning.
What matters
Results depend on how you present the examples.
- Clear, consistent formatting across examples helps the model lock onto the pattern
- Examples that resemble the real query work best
- The prompt has a length limit, so you can only fit so many examples
Key idea
Few shot in context learning teaches a task through examples in the prompt, steering the model without changing any weights.