상황에 맞는 학습

Ordinary fine-tuning updates a model's parameters through training. In-context learning instead puts information and demonstrations into the input supplied at inference. The model's weights remain fixed while its internal activations and predictions depend on that input. A zero-shot prompt gives no task demonstrations, a one-shot prompt gives one, and a few-shot prompt gives several. For example, demonstrations can show support messages paired with the labels refund and billing; a new message can then be presented using the same structure. The examples suggest the desired pattern without guaranteeing a correct classification. The GPT-3 paper evaluated task demonstrations supplied through text without gradient updates and reported improvements with scale across many tasks. That evidence does not establish a universal one-hundred-billion-parameter threshold below which in-context learning is impossible. Prompt design also matters. Lu and colleagues demonstrated sensitivity to the order of few-shot examples in their evaluated classification settings, so reordering identical examples can change results without changing model weights. Demonstrations can illustrate a format or task relationship, but they are still input tokens that must be processed. Adding more examples uses more of the supported context budget and can increase per-request work; it is not automatically a quality improvement. For a stable application, compare prompting and fine-tuning using task quality, training effort, request costs, and maintenance needs instead of assuming one method is always cheaper.