WebRider:面向实时网页辅助的个性化意图控制器
文章背景与核心概要
在处理复杂的网页任务时,仅仅获取最终答案是不够的。真正的自动化代理需要能够执行一套完整的策略,包括明确验证内容、处理不确定性、遵循用户偏好以及确定终止时机。然而,目前的实时网页智能体评估往往只关注最终结果,这掩盖了执行过程中严重的策略违规行为。
为了解决这一问题,研究人员提出了 WebRider 框架。该框架引入了“个性化意图控制器”,将委托策略形式化为“意图契约”,并通过分层架构确保操作的安全性。此外,研究团队还发布了 RiderBench 基准测试,涵盖了 42 个公共网站上的 4,096 个实时网页契约,为评估智能体在任务执行过程中的合规性与一致性提供了重要标准。
📌 执行摘要
委托复杂的网页任务不仅仅是检索最终答案,它要求成功传递一套策略,该策略规定了要验证的内容、如何处理不确定性、哪些偏好至关重要以及何时停止。当前的实时网页智能体通常仅根据其最终答案进行评估,这往往掩盖了过程中严重的策略违规行为。
Delegating complex web tasks requires more than simply retrieving a final answer—it demands the successful transfer of a policy dictating what to verify, how to handle uncertainty, which preferences matter, and when to stop. Current live-web agents are typically evaluated exclusively on their final answers, frequently masking severe policy violations along the way.
为了解决这一差距,作者引入了 WebRider,这是一个为实时网页辅助设计的、具有个性化意图控制器的新型框架。WebRider 将委托策略形式化为意图契约 (intent contracts),利用分层架构安全地执行操作,并引入了 RiderBench,这是一个包含 42 个公共网站上 4,096 个实时网页契约的综合基准测试套件。
To address this gap, the authors introduce WebRider, a novel framework featuring persona-conditioned intent controllers designed for live-web assistance. WebRider formalizes delegated policies into intent contracts, utilizes a hierarchical architecture to execute actions safely, and introduces RiderBench, a comprehensive benchmark suite of 4,096 live-web contracts across 42 public websites.
🔍 关键发现与问题
- 保真度差距 (The Fidelity Gap): 全面的实时审计揭示了智能体性能的鲜明对比:虽然强大的控制器成功完成了 99.2% 的任务,但它仅在 38.8% 的情况下遵守了所有潜在的策略约束。
- 完成 \(\neq\) 保真度: 合理的最终答案很容易掩盖对用户偏好、验证标准和操作约束的潜在违规。
- The Fidelity Gap: A full live audit reveals a stark contrast in agent performance: while a strong controller successfully completes 99.2% of tasks, it honors all underlying policy constraints in only 38.8% of cases.
- Finishing \(\neq\) Fidelity: Plausible final answers can easily conceal underlying violations of user preferences, verification standards, and operational constraints.
🛠️ WebRider 的架构
WebRider 通过将浏览路径视为一等的可审计对象,弥合了任务完成度与策略保真度之间的差距。它依赖于三个核心组件:
WebRider bridges the gap between task completion and policy fidelity by treating the browsing path as a first-class, auditable object. It relies on three core components:
- 意图契约 (Intent Contracts): 一种操作记录,捕获目标、约束、证据义务、答案格式和任务本地的个性化控制,即使在动态网页发生变化时,这些记录依然有效。
- 分层架构 (Hierarchical Architecture):
- 顶层控制器: 维护和管理意图契约。
- 中间层: 将意图转化为受保护的可执行操作。
- 工具层: 在浏览器、搜索和地图界面中执行操作。
- RiderBench 基准测试: 在 42 个公共网站的 4,096 个实时网页契约中评估系统,审计内部契约状态和可见的用户体验,以确保策略的保持和个性的一致性。
- Intent Contracts: An operational record capturing goals, constraints, evidence obligations, answer formats, and task-local persona controls that remain valid even as dynamic web pages change.
- Hierarchical Architecture:
- Top-Layer Controller: Maintains and manages the intent contract.
- Middle Layer: Translates intentions into guarded executable actions.
- Tool Layer: Executes actions across browser, search, and map interfaces.
- RiderBench Benchmark: Evaluates systems across 4,096 live-web contracts on 42 public websites, auditing both internal contract states and the visible user experience for policy preservation and persona consistency.
📄 参考资料与附加链接
- 主要学科: 人工智能 (
cs.AI),人机交互 (cs.HC) - DOI: 10.48550/arXiv.2608.06704
- 访问论文: 查看 PDF | HTML 版本 | TeX 源码
- Primary Subjects: Artificial Intelligence (
cs.AI), Human-Computer Interaction (cs.HC)- DOI: 10.48550/arXiv.2608.06704
- Access Paper: View PDF | HTML Version | TeX Source
