Would it be feasible to do a soft RLHF using steering when an agents gives an undesired response?