LoFiT: Localized Fine-tuning on LLM Representations | Read Paper on Bytez