Small models, big results: Achieving superior intent extraction through decomposition

Google Research · 5 months ago

Researchers developed a two-stage decomposition approach for small multimodal models to extract user intent from sequences of mobile and web interactions entirely on-device. The method separates user interaction understanding into individual screen summarization followed by intent extraction, achieving performance comparable to much larger models like Gemini 1.5 Pro while using Gemini 1.5 Flash 8B at lower cost and faster inference. This approach enables mobile devices to better anticipate user needs and offer contextually relevant suggestions without sending sensitive data to servers.