The REPHRASE Circuit: How Fine-Tuning Enhances LLMs to REPHRASE Text — LessWrong