A technique that combines multiple visual tokens from images into fewer tokens for efficient multimodal model processing.