ParaGUIBench Task Assets (Parallel_benchmark)

Guest-visible input assets and evaluator-side answer files for ParaGUIBench — a 233-task benchmark for studying whether multiple GUI agents can coordinate decomposable workloads across isolated desktop instances and reduce the critical path without weakening task-level evaluation ("Beyond Sequential Interaction: Benchmarking Parallel Execution and Coordination for GUI Agents").

What lives in this repository

Path	Content
benchmark_dataset/<task-uid>/	Guest-visible per-task input files (documents, spreadsheets, presentations, media, archives)
answer_files/<task-uid>/	Evaluator-side expected outputs (never shown to the agent)

The 233 canonical task definitions, evaluation code, orchestration framework, and per-file integrity manifests live in the code repository: https://github.com/pkgunboat/ParaGUIBench. File-level integrity for the generator-managed input assets (49 manifests, 193 files) and the pinned gold entries is enforced by those manifests — repository revision, size, and SHA-256, re-verified on every fetch; remaining answer_files/ directories are fetched at pinned task uids at runtime.

Task provenance (233 tasks)

Source	Tasks	Notes
OSWorld	28	Adapted from upstream OSWorld tasks for parallel multi-VM execution; upstream license Apache-2.0
VeriWeb (originally released as VeriGUI, arXiv:2508.04026)	29	Web-search information-retrieval tasks; upstream license Apache-2.0
Original to ParaGUIBench	176	Includes 91 WebMall online-shopping tasks (executed against four self-hosted WordPress/WooCommerce stores), 57 self-authored file-operation tasks, and 28 further self-authored tasks

Attribution: OSWorld-derived assets originate from xlangai/ubuntu_osworld and xlangai/ubuntu_osworld_file_cache; VeriWeb-derived task content originates from 2077AIDataFoundation/VeriWeb. We thank the authors of both benchmarks. If you are an upstream rights holder and believe any item is misattributed, please open an issue in the code repository.

License

This dataset is released under the Apache License 2.0. Items derived from OSWorld and VeriWeb remain subject to their upstream Apache-2.0 licenses and are redistributed with attribution; the benchmark code and remaining assets are original to ParaGUIBench.

Usage

This dataset is consumed automatically at runtime by the benchmark runners — no manual download is required. To deploy and run the benchmark, follow the code repository's INSTALL.md and the deployment guides under docs/deployment/.

Citation

@software{paraguibench2026,
  title  = {ParaGUIBench: Benchmarking Parallel Execution and Coordination for GUI Agents},
  author = {Yu, Zedong and Li, Qianxing and Gao, Zhi and Xiang, Liuyu and Shi, Chenrui and Liu, Yang and Wu, Huiming and Wei, Yujie and Fei, Yuhao and Fu, Yubo and He, Zhaofeng},
  url    = {https://github.com/pkgunboat/ParaGUIBench},
  year   = {2026}
}

The canonical citation record (including the paper preprint, once public) is maintained in the code repository's CITATION.cff.