Research Papers 论文研究 1d ago Updated 1d ago 更新于 1天前 45

GUI-CC: Benchmarking Contextual Consistency of GUI World Models as Agent Environments GUI-CC:基准测试GUI世界模型作为智能体环境的上下文一致性

GUI-CC is a new benchmark that evaluates contextual consistency of GUI world models as multi-step agent environments, addressing the mismatch between one-step next-screen prediction and real-world agent use The benchmark features two complementary tracks: an offline reference-action track (500 tasks from GUIOdyssey) and an online agent-loop track (200 emulator-verified tasks across 30 mobile apps) GUI-CC evaluates four key metrics: transition fidelity, transition plausibility, contextual consist 提出GUI-CC基准测试,评估GUI世界模型作为多步智能体环境的上下文一致性,而非仅作为单步下一屏预测器 包含两个互补评估轨道:离线参考动作轨道(沿真实移动GUI轨迹运行)和在线智能体循环轨道(固定探测智能体与模型生成UI交互) 构建500个离线轨迹任务(来自GUIOdyssey)和200个模拟器验证的在线任务(覆盖30个移动应用) 评估四个维度:转换保真度、转换合理性、上下文一致性和任务进度 实验发现:合理的单步生成不能保证可靠的环境模拟,当前模型常产生看似可用的界面,但无法保持任务相关上下文或支持可执行的多步展开

58
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • GUI-CC is a new benchmark that evaluates contextual consistency of GUI world models as multi-step agent environments, addressing the mismatch between one-step next-screen prediction and real-world agent use
  • The benchmark features two complementary tracks: an offline reference-action track (500 tasks from GUIOdyssey) and an online agent-loop track (200 emulator-verified tasks across 30 mobile apps)
  • GUI-CC evaluates four key metrics: transition fidelity, transition plausibility, contextual consistency, and task progress
  • Experiments reveal that plausible single-step generation does not guarantee reliable multi-step environment simulation; current models fail to preserve task-relevant context or support executable rollouts

Why It Matters

This benchmark exposes a critical gap in how GUI world models are evaluated versus how they are actually used — as multi-step environments for autonomous agents. For AI practitioners building GUI agents, it highlights that next-screen prediction quality alone is insufficient; contextual consistency across repeated interactions is equally vital for reliable agent performance.

Technical Details

  • Two-track evaluation framework: The offline track rolls models along real mobile GUI trajectories from GUIOdyssey (500 tasks), while the online track deploys fixed probing agents to interact with model-generated UIs across 30 mobile apps (200 emulator-verified tasks)
  • Four evaluation metrics: Transition fidelity (accuracy of state transitions), transition plausibility (visual realism of generated screens), contextual consistency (preservation of task-relevant state across steps), and task progress (ability to advance toward goals over multi-step rollouts)
  • Core finding: Current GUI world models produce visually plausible single-step outputs but degrade significantly when reused iteratively, failing to maintain coherent context or support executable multi-step agent trajectories

Industry Insight

  • Benchmark designers and model developers should shift evaluation focus from isolated next-screen prediction to multi-step consistency, as this better reflects real agent deployment requirements
  • The gap between single-step plausibility and multi-step reliability suggests that training objectives for GUI world models need to incorporate long-horizon consistency losses, not just per-step fidelity
  • For agent system builders, relying on current GUI world models for autonomous multi-step tasks carries significant risk of context drift; validation on benchmarks like GUI-CC should be a prerequisite before production deployment

TL;DR

  • 提出GUI-CC基准测试,评估GUI世界模型作为多步智能体环境的上下文一致性,而非仅作为单步下一屏预测器
  • 包含两个互补评估轨道:离线参考动作轨道(沿真实移动GUI轨迹运行)和在线智能体循环轨道(固定探测智能体与模型生成UI交互)
  • 构建500个离线轨迹任务(来自GUIOdyssey)和200个模拟器验证的在线任务(覆盖30个移动应用)
  • 评估四个维度:转换保真度、转换合理性、上下文一致性和任务进度
  • 实验发现:合理的单步生成不能保证可靠的环境模拟,当前模型常产生看似可用的界面,但无法保持任务相关上下文或支持可执行的多步展开

为什么值得看

本文揭示了GUI世界模型评估中的关键盲区:现有评估过于关注单步预测质量,忽视了多步交互场景下的上下文一致性需求。这对开发GUI智能体的研究者和工程师具有重要参考价值,帮助识别模型在实际应用中的真实能力边界。

技术解析

  • 提出GUI-CC基准测试,专门针对GUI世界模型作为智能体环境时的上下文一致性进行评估,填补了现有评估方法的空白
  • 设计双轨道评估体系:离线参考动作轨道沿真实移动GUI轨迹运行模型,在线智能体循环轨道让固定探测智能体与模型生成的UI进行交互
  • 数据集构建:500个离线轨迹任务来自GUIOdyssey,200个在线任务经模拟器验证,覆盖30个移动应用
  • 评估指标体系包含四个维度:转换保真度(transition fidelity)、转换合理性(transition plausibility)、上下文一致性(contextual consistency)和任务进度(task progress)
  • 实验结论:当前模型在单步生成质量上表现尚可,但在多步环境模拟中存在明显缺陷,无法有效保持任务相关上下文信息

行业启示

  • GUI世界模型的评估范式需要从单步预测转向多步环境模拟,上下文一致性应成为核心评估指标
  • 开发GUI智能体应用时,应重视模型在连续交互中的状态保持能力,而非仅优化单步预测准确率
  • 现有基准测试存在评估盲区,行业需要建立更贴近实际应用场景的多步交互评估标准,以推动GUI世界模型向实用化迈进

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Benchmark 基准测试 Evaluation 评测 Agent Agent Multimodal 多模态 Research 科学研究