Representation Fine-Tuning (ReFT) is a technique to fine-tune LLMs for specific tasks based by only modifying a small fraction of their representations.