CE-Agent: a cloud-edge collaborative agent framework with dynamic offloading and synergistic distillation

Deploying Large Language Model (LLM) agents on edge devices is constrained by hardware memory. Pure cloud architectures resolve this memory limit but incur high latency, consume excessive network bandwidth, and expose sensitive local data. Static Small Language Models (SLMs) at the edge struggle with complex multi-step reasoning tasks. This paper proposes CE-Agent, a cloud-edge collaborative framework. It mitigates the inherent capability-resource dilemma in agentic systems. The framework implements an adaptive, in-process offloading router. It evaluates token generation entropy and local tool execution errors during the multi-step reasoning phase. Upon exceeding a defined uncertainty threshold, the system halts the local execution, compresses the historical textual context, and transfers the remaining inference steps to a cloud LLM. Additionally, the framework establishes a cloud-assisted trajectory distillation loop. The cloud server applies Direct Preference Optimization (DPO) to the successfully offloaded trajectories. It periodically synchronizes Low-Rank Adaptation (LoRA) weights back to the edge. This enables continuous edge capability evolution. Evaluations on the ALFWorld and WebShop datasets demonstrate the system efficacy. CE-Agent achieves a 91.5% task success rate on ALFWorld and 84.2% on WebShop. Compared to pure cloud deployments, it reduces average end-to-end latency by 57.8% and cloud API token consumption by 78.0%. The system remains agnostic to the base model architecture and operates efficiently under 10 Mbps bandwidth constraints. We also discuss limitations regarding privacy, model-collapse risk, and network variability.