Prompting Large Language Models

Imagine you need a model that answers questions about legal documents. You have a clear target, but the number of labeled legal examples you can collect is small. A natural question follows: must the model learn everything about legal language from those few examples alone, or can it start from something it already knows? This article answers that question. The short version is that a model can often start from a foundation that already contains much of what it needs, and the remaining work is mostly refinement.

The idea has a practical motivation. Since the early days of deep learning, transfer learning has offered a way out when the data for a target task is limited or expensive to collect. The strategy is to train a base model on a task where data is abundant, and then move what it learned toward the target task where data is scarce.

People do this constantly without naming it. Someone who already plays the piano usually finds it easier to pick up another instrument, because reading notes, sense of rhythm, and finger coordination carry over. The new skill is not learned from zero. Transfer learning asks the same question of machines: which parts of what was learned for one job still help with a different but related job?

It helps to trace the movement of knowledge step by step. The diagram below lays out the path that a transfer learning workflow follows, from the data-rich starting task to the data-poor destination task. Read it from left to right as a chain of hand-offs.

Abundant data shapes the base model during pre-training. What it learned is kept, not thrown away, and then applied to the new task. The target task contributes only the final adjustment, so the model does not need to rediscover the foundations. The scarce data is spent on refinement rather than on building everything from the beginning.

The relationship between the two terms is one of containment. Transfer learning is the general idea of reusing knowledge. Finetuning is one method that puts the idea into practice. The diagram below shows that finetuning sits inside transfer learning as one method, not as a separate field that competes with it.

Transfer learning improves sample efficiency, which means a model can learn the same behavior from fewer examples. A sample-efficient model learns effectively from a small number of samples. Finetuning brings this benefit directly to specialized work. In the source's framing, knowledge from pre-training on text completion, a task with abundant data, is transferred to more specialized tasks such as legal question answering or text-to-SQL, which often have far less data available.