Reproduce “Automatic Macro Mining from Interaction Traces at Scale” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Forrest Huang, Gang Li, Tao Li, Yang Li
A system leveraging Large Language Models and graph-based optimization automatically extracts semantically meaningful and executable macros from mobile interaction traces at scale. The Google Research team's system generated 23,117 macros from the RICO dataset, achieving a 43.6% reduction in macro length after optimization and a 76.7% execution success rate in live environments.
Key insights
- Large Language Models can effectively bridge the gap between low-level UI elements and high-level user tasks by performing chain-of-thought reasoning for task discovery, action grounding, and parameter identification.
- Graph-based optimization significantly enhances macro efficiency by discovering optimal, shorter execution paths across multiple diverse interaction traces, even if not present in any single trace.
- The system can discover novel and valuable user tasks (48.2% unknown to users in a study) directly from observational data, augmenting existing UI datasets with semantically rich, executable task definitions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.