A new arXiv paper presents a systematic analysis of the components and interactions involved in domain generalization for smartphone-based human activity recognition. The work targets performance drops that occur when models face shifts in users, devices, sensor placement, environments, or data collection protocols.
A new arXiv technical report introduces BlueLM-GUI, a mobile GUI agent trained end-to-end on real devices rather than in simulated sandboxes. The authors argue that sandbox-based training creates a mismatch with production conditions and propose a self-improving "flywheel" loop that collects and reuses real-device interaction data. It targets the practical obstacles industrial teams face when deploying mobile agents, including data collection cost and device-specific behavior.
Calif Research published a demonstration of a worm it calls WeWorm, which it says spreads through WeChat voice calls on both iOS and Android without any action from the target. According to the researchers, the payload can propagate even when the recipient never picks up the call or touches their device. The work is presented as a demo rather than evidence of an active campaign in the wild.