Steering vectors align LLMs with human values

Linear directions extracted from large‑language‑model activation distributions map onto human‑value...