Researchers just built a benchmark to test whether mobile AI agents can guess what you want before you ask.
A new paper introduces Act2Intention, a framework and benchmark for training phone-based AI agents to predict user intent from screen actions instead of just executing typed commands. The team built the Act2Intention Bench, a dataset of 72,511 recorded intentions and more than 700,000 individual actions collected across 52 apps. They then trained the Act2Intention Agent in three stages: reading a user's intent from what is on screen, predicting what that user will want to do next, and executing decisions based on past experience. Fine-tuning on the benchmark produced measurable gains over an untuned baseline using the same agent setup, with improvements of 32.0 points on an intention-understanding accuracy metric, 10.25 points on prediction accuracy, and 6.9 points on an execution success metric.
Most mobile AI agents today are reactive. They wait for a command, then tap through screens to complete it. This work pushes toward agents that anticipate needs, the same direction predictive suggestion features from the major phone makers have nudged toward for years without ever publishing how well they actually work. A public benchmark with three quarters of a million logged actions gives other labs a common yardstick, instead of every team claiming proactive behavior on its own terms.
The gains look strong, but they are measured on the team's own benchmark, and teaching software to predict your next tap from your usage history is also a tidy euphemism for building a behavioral profile. That trade-off is worth remembering before proactive agents start shipping on phones.