Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks