Administrative Decentralization in Edge-Cloud Multi-Agent for Mobile Automation
New edge-cloud AI architecture cuts cloud token use by 37.5% and boosts task success by 21.7%.
A team of researchers, including Senyao Li and Zhigang Zuo, has published a paper on arXiv introducing AdecPilot, a novel framework applying administrative decentralization to edge-cloud multi-agent systems for mobile automation. The core innovation is decoupling high-level strategic design, handled by a UI-agnostic cloud agent, from low-level tactical grounding and execution, managed by an autonomous bimodal team on the edge device. This shift addresses the 'cognitive lag' of centralized systems where the cloud makes all decisions, by empowering the edge to plan and self-correct locally.
AdecPilot employs a Hierarchical Implicit Termination (HIT) protocol to enforce deterministic stops and prevent post-completion hallucinations—a common failure mode in AI agents. Extensive experiments demonstrate significant performance gains: the framework improves task success rate by 21.7% compared to EcoAgent, reduces end-to-end latency by a dramatic 88.9% against the CORE framework, and cuts cloud token consumption by 37.5%. This architecture makes mobile automation—like controlling apps or performing sequences of actions on a phone—much faster, more reliable, and less dependent on constant cloud connectivity.
- Decouples strategy (cloud) from tactics (edge), enabling local planning and self-correction without cloud calls.
- Uses a Hierarchical Implicit Termination protocol to prevent post-completion hallucinations and ensure deterministic stopping.
- Achieves an 88.9% reduction in end-to-end latency and a 37.5% cut in cloud token usage versus prior systems.
Why It Matters
Enables faster, more reliable, and private on-device AI automation for mobile apps, reducing cloud dependency and cost.