ContextWeave
收藏资源简介:
ContextWeave是由复旦大学、上海创新研究院和字节跳动联合构建的纵向工作流基准数据集,旨在评估语言代理在真实办公场景中利用记忆提升下游任务性能的能力。该数据集包含从14位参与者数月工作流中重建的1,005个可执行任务,其中568个为核心评估任务,涵盖指令、容器化环境、轨迹及评分细则。通过严格的隐私脱敏、任务重构和可控轨迹对齐流程,数据集保留了真实工作流的时序依赖和用户偏好。该基准聚焦于衡量记忆对工作空间质量和个性化偏好对齐的改进效果,并诊断相关性、连续性、可解性和误导鲁棒性,旨在推动记忆系统超越单纯检索,实现可靠的经验复用与工作流延续。
ContextWeave is a longitudinal workflow benchmark dataset jointly constructed by Fudan University, Shanghai Institute of Innovation and ByteDance, which aims to evaluate the capability of language agents to leverage memory to enhance the performance of downstream tasks in real-world office scenarios. This dataset includes 1,005 executable tasks reconstructed from the multi-month workflows of 14 participants, among which 568 are core evaluation tasks covering instructions, containerized environments, execution trajectories and scoring criteria. Through strict privacy de-identification, task reconstruction and controllable trajectory alignment procedures, the dataset retains the temporal dependencies and user preferences of real-world workflows. This benchmark focuses on measuring the improvement effect of memory on workspace quality and personalized preference alignment, as well as diagnosing correlation, continuity, solvability and robustness against misleading information, with the goal of advancing memory systems beyond simple retrieval to achieve reliable experience reuse and workflow continuity.
ContextWeave 数据集详情
概述
ContextWeave 是一个用于评估编码智能体在长期个人工作日志任务中记忆系统的基准测试集。每个任务提供参考工作区状态、参考消息历史、任务元数据和评估规则。在智能体运行前注入记忆组件,基准测试会比较有记忆与无记忆基线的行为差异,并与参考记录进行对比。
数据集内容
数据组成
data/tasks/:包含任务元数据、参考轨迹、资源和图像引用data/meetings/:包含选定任务使用的会议和上下文资源data/metrics/:包含偏好规则和参考历史摘录,其中:preference/:按人员分类的全局偏好规则relevance/:按人员分类的参考历史摘录
数据格式
对于每个人员,data/tasks/<person>/ 包含:
workflow.json:子任务起始镜像信息message_logs/:每个子任务的参考消息日志(JSON 格式)task/:任务说明、注释和参考文档变更_resources/:子任务使用的文件和中间件资源
data/metrics/preference/<person>.json 包含全局偏好规则;data/metrics/relevance/<person>.json 包含语义相关性评估所需的任务 ID 和摘录文本。
数据压缩包
基准数据以压缩归档文件形式提供:data/task_data.tar.zst,需通过 tar --zstd -xf task_data.tar.zst 命令解压后使用。
论文信息
该数据集出自 arXiv 论文 “ContextWeave: A Real-World Workflow Benchmark”(arXiv:2608.04830),作者包括 Wang Bo、Yao Yuqian、Wang Enxi 等,由复旦大学、上海创新研究院和字节跳动联合开发。
引用方式
如需引用 ContextWeave,建议使用以下 BibTeX 格式:
bibtex @article{wang2026contextweave, title={ContextWeave: A Real-World Workflow Benchmark}, author={Wang, Bo and Yao, Yuqian and Wang, Enxi and Jin, Luozhijie and Liu, Yang and Suo, Yiran and Cai, Yuxuan and Zhou, Enyu and Gao, Yufei and Guo, Honglin and Huai, Tianyu and Ji, Li and Lei, Zhikai and Li, Bufan and Lin, Lizhi and Liu, Jinxiu and Yang, Jie and Zhou, Jiazheng and Zhou, Maosen and Qian, Pengfang and Liu, Shichun and Liu, Guanshan and Zheng, Hao and Yu, Yunhao and Yan, Hang and Kang, Jihua and Chen, Xinchi and Qiu, Xipeng}, journal={arXiv preprint arXiv:2608.04830}, year={2026} }





