跳转至

文章背景与核心概要

当前的AI智能体在实现个性化时往往面临瓶颈,主要原因在于它们将用户视为“扁平化的档案”。当用户在不同生活场景(如工作与个人生活)中存在冲突的优先级时,这种方法往往会失效。本文介绍了 SIMGUIDE,这是一种创新方法,它将用户上下文组织成结构化的、特定于领域的模块,称为“Sims”(模拟模块)。

通过将这些约束条件与来自过去决策的程序化示例相结合,SIMGUIDE 在复杂的规划任务中显著优于标准的基于检索的个性化方法(RAG)。该研究不仅提出了结构化表征方案和全新的诊断测试集 SIMBENCH,还在多个主流模型上验证了其在偏好遵循和任务规划方面的显著性能提升。


SIMGUIDE:程序化落地的多上下文表示,赋能个性化智能体规划 (SIMGUIDE: Procedurally Grounded Multi-Context Representations for Personalized Agent Planning)

作者: Chirag Shah
日期: 2026年5月14日
学科: 人工智能 (cs.AI)
arXiv ID: 2608.24888


执行摘要 (Executive Summary)

当前的AI智能体在实现个性化时往往面临瓶颈,主要原因在于它们将用户视为“扁平化的档案”。当用户在不同生活场景(如工作与个人生活)中存在冲突的优先级时,这种方法往往会失效。本文介绍了 SIMGUIDE,这是一种创新方法,它将用户上下文组织成结构化的、特定于领域的模块,称为“Sims”。通过将这些约束条件与来自过去决策的程序化示例相结合,SIMGUIDE 在复杂的规划任务中显著优于标准的基于检索的个性化方法(RAG)。

Executive Summary

Current AI agents struggle with personalization because they treat users as "flat profiles." This approach fails when users have conflicting priorities across different life contexts (e.g., work vs. personal life). This paper introduces SIMGUIDE, a novel method that organizes user context into structured, domain-specific blocks called "Sims." By grounding these constraints with procedural examples from past decisions, SIMGUIDE significantly outperforms standard retrieval-based personalization (RAG) in complex planning tasks.


核心贡献 (Key Contributions)

1. “Sim” 表征方法

SIMGUIDE 没有将扁平的用户档案直接拼接到提示词中,而是将上下文结构化为 Sims: * 类型化、特定领域的模块: 上下文按领域进行分段。 * 程序化落地: 每个约束条件都与用户历史中的真实程序化示例相链接,从而使智能体不仅能理解偏好是什么,还能理解如何采取行动。

1. The "Sim" Representation

Instead of concatenating a flat user profile into a prompt, SIMGUIDE structures context into Sims: * Typed, Domain-Specific Blocks: Context is segmented by domain. * Procedural Grounding: Each constraint is linked to real-world procedural examples from the user's history, allowing the agent to understand how to act rather than just what the preferences are.

2. SIMBENCH:全新的诊断测试集

作者推出了 SIMBENCH,这是一个包含 47 个偏好条件规划任务的基准测试。这是第一个旨在测试智能体根据活跃用户上下文切换行为能力的基准测试——这也是实现真正个性化AI的关键要求。

2. SIMBENCH: A New Diagnostic Suite

The author introduces SIMBENCH, a benchmark consisting of 47 preference-conditioned planning tasks. This is the first benchmark designed to test an agent's ability to switch behaviors based on active user context—a critical requirement for truly personalized AI.

3. 性能提升

  • 程序化落地对比 RAG: 在 GPT-4o 上,经过程序化落地的 Sims 比标准 RAG 的偏好遵循度高出 +7.9 个百分点\(p = 0.013\))。
  • 跨模型鲁棒性: 这些结果在 GPT-4o 和 Claude Sonnet 4.5 的 100 个 \(\tau\)-bench 任务中得到了复现(\(p \leq 0.023\))。
  • 参数化自适应: 任务匹配的 LoRA 微调将生成质量提高了 12.8 个 ROUGE-L 点。按“Sim 类型”而非用户身份进行路由适配器选择,额外带来了 7.3 点的提升,这证明了表征的格式比内容本身更具决定性的设计变量。

3. Performance Gains

  • Procedural Grounding vs. RAG: On GPT-4o, procedurally grounded Sims outperformed standard RAG by +7.9 Preference Adherence points (\(p = 0.013\)).
  • Cross-Model Robustness: These results were replicated across both GPT-4o and Claude Sonnet 4.5 on 100 \(\tau\)-bench tasks (\(p \leq 0.023\)).
  • Parametric Adaptation: Task-matched LoRA fine-tuning improved generation quality by 12.8 ROUGE-L points. Routing adapters by "Sim type" rather than user identity provided an additional 7.3-point boost, demonstrating that the format of the representation is a more critical design variable than the content itself.

访问与资源 (Access & Resources)

Access & Resources


许可协议 (License)

license icon 查看许可协议

License

license icon View License