Knowledge-driven Reasoning for Mobile Agentic AI: Concepts, Approaches, and Directions

A research team from Nanyang Technological University, including Guangyuan Liu and Dusit Niyato, has published a new framework titled 'Knowledge-driven Reasoning for Mobile Agentic AI.' The paper addresses a critical bottleneck: running sophisticated AI models on resource-constrained mobile platforms like edge robots and drones, which face strict size, weight, power, and cost (SWAP-C) limits and intermittent connectivity. Instead of just optimizing single communication rounds, their framework focuses on sustaining AI competence across a continuous stream of tasks by extracting and reusing knowledge from past executions.

Their core innovation is a DIKW-inspired taxonomy that classifies knowledge into four representations—retrieval, structured, procedural, and parametric—each offering a distinct trade-off between reasoning speed and failure risk. The key insight is that 'knowledge exposure is non-monotonic'; too little forces costly trial-and-error, while too much introduces conflicting cues. The framework compresses learned decision structures into compact 'knowledge packs' that can be synchronized over bandwidth-limited links and injected into the on-device model's reasoning process.

In a practical UAV case study, the framework validated its effectiveness. By synchronizing a small knowledge pack over an intermittent backhaul link, the researchers enabled a 3-billion-parameter model running directly on the drone to achieve perfect mission reliability. This on-device, knowledge-augmented reasoning demonstrated lower latency, energy consumption, and error accumulation compared to both 'knowledge-free' on-device reasoning and cloud-centric replanning approaches. This demonstrates a viable path to deploying more capable AI agents in real-world, offline-first environments.