T: Localized Fine-tuning on LLM Representations