Papers on mechanistic interpratability and representation engineering, e.g. from Anthropic would be a good start.