EC²-Bench Evaluates Edge-Cloud Collaboration for Long-Horizon AI Agents
Northeastern University, Tsinghua University and China Mobile researchers released EC²-Bench, an open benchmark for edge-cloud collaboration in long-horizon AI agent tasks. It contains 128 executable digital-workspace tasks across six scenarios, with 100 tasks carrying fine-grained privacy annotations covering 3,385 sensitive information units, and evaluates task performance, resource costs and privacy exposure. The study compares four collaboration strategies; with Qwen3.5-9B, the best strategy raised the completion score from 39.12 under edge-only execution to 60.46 while using cloud resources costing 15% of the cloud-only setup.
EC²-Bench offers a concrete open framework for measuring whether edge-cloud agent designs balance task success, cloud spending and privacy exposure throughout multi-step workflows.
10月26日消息，近日，由东北大学、清华大学、中国移动的研究团队共同发布EC²‑Bench（Edge‑Cloud Collaborative Agent Benchmark），一套面向大模型智能体长周期任务的端‑云协同评测基准，系统研究边缘设备与云端大模型如何分工协作完成复杂智能体任务，同时兼顾任务性能、算力成本与隐私泄露风险，相关代码、数据集、论文已全部开源开放。
图 ：总览。一个典型的长程工作流里，端侧和云端 Agent 在同一条轨迹上轮流执行；我们研究的四种协同策略，区别在于谁决定云端何时介入、云端介入后做什么；以及三个研究问题。
当前本地边缘大模型（通常 10B 参数以内）数据不出设备、隐私性好，但面对复杂规划、工具调用类长周期任务能力存在明显天花板；云端前沿大模型能力强，但调用成本高昂，且会接收本地敏感工作数据，带来隐私泄露隐患。
过去 RouterBench、M4A1 等评测基准大多聚焦单次查询的路由选择，忽略智能体多步骤完整执行轨迹中逐步累积的成本、隐私泄露问题，缺少面向完整任务流程的端云协同评测体系。
#EC²-Bench#Northeastern University#Tsinghua University#China Mobile#OpenClaw#Qwen3.5-9B#Qwen3.5-2B#Qwen3.5-27B#Qwen3.7-Plus#GLM-5.3#DeepSeek-V4.1-Flash