In-context learning (ICL) is when a language model adapts its behavior based on examples or instructions placed directly in its prompt — with no weight updates, no retraining, and no gradient steps. You simply show the model a few examples of what you want, and it generalizes from them on the fly.

This is fundamentally different from training, where a model's internal parameters are adjusted by processing massive datasets over many compute-intensive steps. Training is expensive, slow, and permanent. In-context learning is instant, cheap, and temporary — the model "forgets" the examples the moment the conversation ends.

The surprising thing is how well it works. A model shown three examples of a translation task will often outperform a smaller model fine-tuned specifically for that task. Researchers are still working out why this happens, but it's one of the most practically useful — and theoretically puzzling — behaviors to emerge from large-scale language model training.