SimGuide：程序化多上下文表示实现个性化代理规划
原标题：SIMGUIDE: Procedurally Grounded Multi-Context Representations for Personalized Agent Planning
AI 摘要
本文提出 SimGuide 方法，将用户上下文结构化为类型化、领域特定的 Sims 块，并用程序化示例进行约束，以解决个性化 AI 代理将用户视为单一实体而导致的冲突问题。作者构建了 SimBench 基准（47 个任务），实验表明程序化约束的 Sims 在 GPT-4o 上比 RAG 方法提升 7.9 个偏好遵循点，并在 Claude Sonnet 4.5 上得到验证。此外，基于 Sim 类型的 LoRA 微调比基于用户身份的适配额外提升 7.3 ROUGE-L 点，表明表示格式是首要设计变量。
正文节选
SimGuide: Procedurally Grounded Multi-Context Representations for Personalized Agent Planning Abstract Personalized AI agents overwhelmingly treat users as single entities: a flat profile concatenated into a prompt. This fails when the same person holds different priorities across life contexts—and fails catastrophically when those priorities conflict. The core problem is not that agents lack information about users; it is that the format of user representations determines whether an agent can a