Datasets:

chengpingan
/

EC2-Bench

license: mit
pretty_name: EC²-Bench (EC2-Bench)
language:
  - en
  - zh
size_categories:
  - n<1K
task_categories:
  - text-generation
tags:
  - agents
  - llm-agents
  - benchmark
  - edge-cloud-collaboration
  - edge-ai
  - on-device
  - llm-routing
  - privacy
  - openclaw
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/tasks.jsonl

EC²-Bench (EC2-Bench)

Empowering Edge Agents: A Systematic Analysis of Edge-Cloud Collaboration in LLM Agent Execution

📄 Paper (arXiv 2610.17816) · 💻 Code (GitHub) · 🌐 Project Page

EC2-Bench is a benchmark for edge-cloud collaboration in long-horizon LLM agents, jointly measuring task performance, resource usage and privacy exposure.

Edge-side models keep data local and cost little, but are weak at demanding planning and tool use. Cloud-hosted frontier models are far more capable, but expensive, and they see everything they are sent. In long-horizon agent execution, cloud participation is no longer a one-shot routing decision per query: it is interleaved throughout the trajectory, so every cloud call shapes later steps, accumulates cost, and decides how much workspace information leaves the device.

What is in this dataset

Path	What it is
`data/tasks.jsonl`	Index of all 128 tasks (shown in the viewer above): prompt, expected behavior, grading criteria, category, source benchmark, language, and the number of annotated sensitivity units
`workspace/ACE_Bench.tar.gz`	All task workspaces in one tarball (files, mock services, and privacy ground truth under `gt/sensitive_items.json`)
`workspace/ACE_Bench/`	The same workspaces, unpacked, for browsing
`Images/acebench-openclaw-v1.0.tar.gz`	Prebuilt Docker image (`docker save` tarball) bundling the OpenClaw harness and per-task mock services

The executable task definitions (with inline grade() verifiers) and the evaluation code live in the GitHub repo. Task IDs keep the benchmark's earlier working name (ACE_Bench).

At a glance

128 executable digital-workspace tasks in 6 categories (office & daily tasks, information search, safety & security, development & operations, data analysis, automation), in English and Chinese.
100 tasks with fine-grained privacy annotations: 3,385 sensitivity units covering personal and organizational information.
Tasks are adapted from Claw-Eval, WildClawBench, QwenClawBench, LiveClawBench, PinchBench and ClawBench, plus 7 newly curated tasks, and standardized to one OpenClaw runtime.

Evaluation axes

Task performance: weighted completion score (rule-based checks, environment-state auditing, rubric-based LLM judge) and Pass³.
Resource usage: cloud input / cache-read / output tokens with USD cost, plus edge-side FLOPs.
Privacy exposure: fraction of annotated sensitivity units that become visible to the cloud anywhere along the trajectory (lower is better).

Four collaboration strategies are studied: Query Routing, Per-step Routing, Edge-controlled Delegation and Edge-controlled Advisory, against edge-only and cloud-only baselines.

Quick start

# task workspaces
hf download chengpingan/EC2-Bench workspace/ACE_Bench.tar.gz --repo-type dataset --local-dir .
tar -xzf workspace/ACE_Bench.tar.gz

# prebuilt Docker image (Hugging Face is not a Docker registry: download, then docker load)
hf download chengpingan/EC2-Bench Images/acebench-openclaw-v1.0.tar.gz --repo-type dataset --local-dir .
docker load -i Images/acebench-openclaw-v1.0.tar.gz

from datasets import load_dataset
tasks = load_dataset("chengpingan/EC2-Bench", split="test")

See the GitHub README for running edge-only, cloud-only and collaborative agents and for grading.

Key findings

Collaboration works for long-horizon agents. Partial cloud participation improves on edge-only execution at a fraction of cloud-only cost and exposure.
Stronger edges close the performance gap, not the privacy gap. As the edge model gets stronger, completion differences across strategies shrink, while privacy exposure stays strongly strategy-dependent.
The best collaboration mode shifts with edge capability, from external routing and direct cloud execution toward edge-controlled invocation and cloud advisory. Deciding when to invoke the cloud and following cloud guidance are the two edge-side abilities that matter.

Full results: paper and leaderboard.

Citation

@article{huang2026ec2bench,
  title   = {Empowering Edge Agents: A Systematic Analysis of Edge-Cloud Collaboration in LLM Agent Execution},
  author  = {Huang, Pengcheng and Liu, Zhenghao and Ren, Shixin and Yu, Shi and Yao, Sijia and Yan, Yukun and Xiao, Chaojun and Sun, Maosong},
  journal = {arXiv preprint arXiv:2610.17816},
  year    = {2026}
}

License

MIT, same as the code repository. Adapted tasks are used under the licenses of their source benchmarks: Claw-Eval, WildClawBench, QwenClawBench, LiveClawBench and PinchBench are MIT-licensed, and ClawBench is Apache-2.0.