Investigating task-specific prompts and sparse autoencoders for activation monitoring

Open in new window