[论文] Self-Evolving World Models for LLM Agent Planning

论文概要

研究领域: Agent 作者: Xuan Zhang, Wenxuan Zhang, See-Kiong Ng 发布时间: 2026-07-01 arXiv: 2507.00003

中文摘要

世界模型为长程LLM智能体提供了一种系统性的方式赋予其预见能力：在执行前预测动作后果。然而，不可靠的预见可能被忽略、误用，甚至降低下游决策质量。本文介绍WorldEvolver，一个自进化的世界模型框架，它在保持下游智能体和所有模型参数冻结的同时，修订其部署时的上下文。WorldEvolver集成了三个模块：(i) 情景记忆，通过基于检索的模拟利用真实动作转换；(ii) 语义记忆，从预测-观测不匹配中提取持久启发式规则；(iii) 选择性预见，在将预测整合到智能体推理上下文之前过滤低置信度预测。我们在ALFWorld和ScienceWorld上评估WorldEvolver，在Word2World上测量世界模型预测准确性，在AgentBoard上测量下游智能体成功率。大量实验表明，WorldEvolver在三个骨干模型上实现了最高的预测准确性，并在下游智能体成功率上领先其他世界模型基线，证明测试时记忆修订同时增强了预测保真度和规划性能。

原文摘要

World models offer a principled way to equip long-horizon LLM agents with foresight: predictions of action consequences before execution. However, unreliable foresight can be ignored, misused, or even degrade downstream decision-making. In this paper, we introduce WorldEvolver, a self-evolving world model framework that revises its deployment-time context while keeping the downstream agent and all model parameters frozen. WorldEvolver integrates three modules: (i) Episodic Memory, which exploits real action transitions through retrieval-based simulation; (ii) Semantic Memory, which extracts persistent heuristic rules from prediction-observation mismatches; and (iii) Selective Foresight, which filters low-confidence predictions before integrating them into agent reasoning context. We evalua...

--- *自动采集于 2026-07-01*

#论文 #arXiv #Agent #小凯