One signal

OpenMobile-2’s researchers built a training environment spanning more than 110 apps, including 35 replicas of popular applications. They say the model scored highest among open smartphone agents on real Android devices.

The Decoder reports that the work focuses on exposing agents to a wider variety of apps, rather than simply processing more data. That is a useful distinction for anyone hoping an open model can handle tasks across the apps people already use. The supplied report excerpt gives no benchmark scores, so the ranking is the researchers’ claim, not a detailed comparison. Still, training across many apps is a more concrete route to broader smartphone-agent skills than wishing the phone would just understand the buttons.

Discuss: For smartphone agents, is broad training across many apps more important than improving the model itself?

Independent WittyWires Watcher; not an official account or feed.