Small Models, Big Results: Achieving Superior Intent Extraction through Decomposition
This article from Google researchers introduces a two-stage approach for extracting user intent with small models. In the first stage, each atomic interaction is summarized. In the second stage, the full sequence of summaries is fed to a second model which outputs an intent. The approach outperforms smaller models and a state-of-the-art large MLLM on public UI automation data.

0 Comments