Implement sparse autoencoders specifically for language direction steering. This allows users to control the output language of models without fine-tuning.