返回全部动态
因果世界模型何时帮助模块化 LLM 智能体
原标题：When Do Causal World Models Help Modular LLM Agents
AI 摘要
该论文研究模块化 LLM 智能体中的因果世界模型何时真正有用。作者提出 FedCausalCompose 框架，通过局部动作提供干预-响应证据来恢复跨模块因果接口，并从理论上证明观测式世界模型在未阻断后门路径下存在不可消除的干预误差。实验表明，因果接口在结构化工具环境中帮助最大，而对话和叙事环境通常需要注意力锚点才能让因果信息影响决策。
以上摘要由 AI 生成，可能存在误差。事实请以原文为准。
正文节选
When Do Causal World Models Help Modular LLM Agents Abstract LLM agents increasingly act through modular systems, such as order, payment, inventory, and shipment services, where actions in one module change which transitions are valid in another. Standard world models usually fit observational traces, but this is not the quantity needed for intervention-time planning: a trace may show that payment precedes shipment without identifying whether payment authorizes shipment, inventory mediates the e
发布时间：2026-10-02 12:00
抓取时间：2026-10-02 12:01
来源机构：arXiv