文章背景与核心概要
当前的电脑操作智能体(Computer-Use Agents)评估方法呈现两极分化状态,主要集中在长周期工作流基准测试和原子级 GUI 接地测试(GUI-grounding tests)之间。ComponentBench 正是为了填补这一“缺失的中层(missing middle)”而提出的——它专注于真实、以组件为中心的交互,既具备足够的复杂度以反映现代 UI 的挑战,又足够精简以支持精准的诊断分析。
研究人员开发了一个与框架无关的包含 97 个规范 UI 组件的本体库(ontology),并据此构建了 2,910 个经过程序验证的任务。通过对七个主流模型(包括 GPT-5.4、Gemini 3 和 Qwen3-VL)的评估,该研究表明,观测与动作空间的设计选择可以使任务成功率发生超过 30% 的波动,这凸显了当前智能体在感知和交互数字界面时存在的显著瓶颈。
ComponentBench: Diagnosing Component-Level Failures in Computer-Use Agents
Authors: Tianchen Guan, Xinlei Lin, Royce Cheng-Yue, Xiangjun Wang, Shuyan Zhou
Published: August 18, 2026 (Accepted at COLM 2026)
arXiv ID: 2608.18307
Summary
Current evaluation methods for computer-use agents are polarized between long-horizon workflow benchmarks and atomic GUI-grounding tests. ComponentBench addresses the "missing middle"—the need for realistic, component-centered interactions that are complex enough to reflect modern UI challenges but concise enough to allow for precise diagnostic analysis.
The researchers developed a library-agnostic ontology of 97 canonical UI components, resulting in 2,910 programmatically verified tasks. By evaluating seven major models (including GPT-5.4, Gemini 3, and Qwen3-VL), the study reveals that observation and action space design choices can shift task success rates by over 30%, highlighting significant bottlenecks in how current agents perceive and interact with digital interfaces.
当前的电脑操作智能体(Computer-Use Agents)评估方法呈现两极分化状态,主要集中在长周期工作流基准测试和原子级 GUI 接地测试(GUI-grounding tests)之间。ComponentBench 正是为了填补这一“缺失的中层(missing middle)”而提出的——它专注于真实、以组件为中心的交互,既具备足够的复杂度以反映现代 UI 的挑战,又足够精简以支持精准的诊断分析。
研究人员开发了一个与框架无关的包含 97 个规范 UI 组件的本体库(ontology),并据此构建了 2,910 个经过程序验证的任务。通过对七个主流模型(包括 GPT-5.4、Gemini 3 和 Qwen3-VL)的评估,该研究表明,观测与动作空间的设计选择可以使任务成功率发生超过 30% 的波动,这凸显了当前智能体在感知和交互数字界面时存在的显著瓶颈。
Key Features of ComponentBench
- Comprehensive Ontology: 97 canonical UI components across widely used libraries.
- Scalable Diagnostic Pipeline: Enables auditing of structural difficulty and synthesis of failure analyses across different component families.
- Human-Centric Benchmarking: Includes cleaned human reference trajectories to evaluate both task success and interaction efficiency.
- Performance Insights: Demonstrates that even the most advanced models take 3.7x longer than humans to complete tasks, with spatial manipulation remaining a persistent challenge.
ComponentBench 的核心特性
- 全面的本体库: 涵盖主流UI库中的 97 个规范 UI 组件。
- 可扩展的诊断流水线: 能够审查结构难度,并综合分析不同组件系列的故障。
- 以人为本的基准测试: 包含清洗过的人类参考轨迹,用于评估任务成功率和交互效率。
- 性能洞察: 表明即使是最先进的模型,完成任务所需的时间也是人类的 3.7 倍,空间操作仍然是一个长期的挑战。
Model Performance & Observations
The study evaluated seven models across four distinct observation and action spaces. A critical finding is the sensitivity of model performance to input representation: * GPT-5 mini: Achieved 83.1% success using accessibility-tree observations, which plummeted to 48.9% when restricted to coordinate-only pixel control. * Efficiency Gap: Even the most optimized configurations significantly lag behind human performance, requiring nearly 4x the time to complete identical tasks.
模型性能与观察
该研究在四种不同的观测和动作空间下评估了七个模型。一个关键发现是模型性能对输入表示形式高度敏感: * GPT-5 mini: 在使用无障碍树(accessibility-tree)观测时取得了 83.1% 的成功率,而当受限于纯坐标像素控制时,成功率骤降至 48.9%。 * 效率差距: 即使是最优化的配置,也明显落后于人类表现,完成相同任务所需的时间接近人类的 4 倍。
Resources
- Website: componentbench.com
- Code: GitHub Repository
- Data: Hugging Face Dataset
相关资源
- 官方网站: componentbench.com
- 代码仓库: GitHub Repository
- 数据集: Hugging Face Dataset
Metadata
- Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Human-Computer Interaction (cs.HC)
- DOI: 10.48550/arXiv.2608.18307
- Format: 30 pages (10 pages main text), 10 figures, 15 tables.
元数据
- 学科分类: 人工智能 (cs.AI);计算与语言 (cs.CL);人机交互 (cs.HC)
- DOI: 10.48550/arXiv.2608.18307
- 篇幅格式: 30 页(10 页正文),10 张图表,15 个表格。