type: concept tags: [edge-computing, offloading, llm, mobile, world-model] related: [[edge-cloud-offloading]], [[on-device-inference-memory-pressure]], [[edgeflow-cold-start]] sources: - url: https://arxiv.org/abs/2602.13628v1 title: "Compact LLM Deployment and World Model Assisted Offloading in Mobile Edge Computing" date: 2026-02 created: 2026-04-14
紧凑 LLM 部署与世界模型辅助的移动边缘卸载¶
概述¶
这篇论文提出了一种结合紧凑型 LLM 部署和世界模型辅助决策的移动边缘计算卸载策略。
核心思路¶
- 紧凑模型部署:在边缘节点部署小型 LLM(1-3B)
- 世界模型预测:用轻量级世界模型预测任务复杂度
- 智能卸载:基于预测结果决定在本地还是云端执行
为什么重要¶
这是 [[edge-cloud-offloading]] 架构的重要进展。传统卸载策略基于规则或简单启发式,而世界模型辅助方法可以更准确地预判任务需求,减少不必要的网络往返。与 [[sustainability-ondevice-intelligence]] 的能耗分析结合,可以实现真正的能效最优调度。
核心问题¶
We prove an Alexandrov-Bakelman-Pucci type estimate, which involves the integral of the determinant of the complex Hessian over a certain subset. It improves the classical ABP estimate adapted (by inequality $2^{2n}|\det(u_{i\bar{j}})|^2\geq |\det(\nabla^2u)|$) to complex setting. We give an application of it to derive sharp gradient estimates for complex Monge-Ampère equations. The approach is based on the De Giorgi iteration method developed by Guo-Phong-Tong for equations of complex Monge-Ampère type.
为什么重要¶
本研究/产品对手机端 AIOS 生态有重要参考价值。推动端侧 AI 从概念走向实际部署。
关联¶
- [[edge-cloud-offloading]] — 端云协同整体架构
- [[mnn-350]] — 边缘推理框架
- [[edgeflow-cold-start]] — 本地模型启动优化