Sparse Autoencoders can steer language models better than thought

Original: Steering LLMs? Actually, Sparse Autoencoders can outperform simple baselines

Source: arxiv.org ↗