Brief

Researchers have introduced H-RePlan, a novel hierarchical replanning framework designed to enhance the robustness of multi-device agent systems. This framework addresses limitations in current systems by enabling agents to distinguish between device-local failures that can be repaired and those requiring broader replanning. To assess its effectiveness, a new fault-injected benchmark called HeraBench was developed, which simulates cross-device workflows on Linux and Android devices. Experiments demonstrate that H-RePlan significantly improves task completion rates and reduces token costs compared to existing baselines. AI