Agentic Environment Engineering for Large Language Models: A Survey of Environment Modeling, Synthesis, Evaluation, and Applications
English summary
This paper systematically surveys the engineering of specialized environments for training and evaluating large language model (LLM) agents. It categorizes environments by lifecycle stages and paradigms, distinguishing between symbolic and neural approaches to environment modeling and automated synthesis. The survey reviews evaluation methods for these environments and outlines evolution paths, including neural-driven, difficulty-driven, and scaling-driven strategies. Future directions such as Environment-as-a-Service, multi-agent environments, and neural-symbolic integration are discussed.
Chinese summary
该论文系统综述了用于训练和评估大语言模型智能体（LLM agents）的专用环境工程。研究根据生命周期阶段和范式对环境进行分类，区分了基于符号和神经方法的环境建模与自动合成。综述回顾了环境的评估方法，并梳理了神经驱动、难度驱动和规模驱动等环境演化路径。未来方向包括环境即服务、多智能体环境和神经-符号集成。
Key points
Proposes a taxonomy of agentic environments across modeling, synthesis, evaluation, and application stages.
提出涵盖建模、合成、评估与应用阶段的智能体环境分类体系。
Identifies symbolic and neural paradigms for environment modeling and automated synthesis.
识别了用于环境建模与自动合成的符号范式和神经范式。
Reviews diverse evaluation methods and environment evolution strategies (neural-driven, difficulty-driven, scaling-driven).
回顾了多种评估方法以及环境演化策略（神经驱动、难度驱动、规模驱动）。
Outlines future directions, including Environment-as-a-Service, multi-agent environments, and neural-symbolic integration.
展望了环境即服务、多智能体环境和神经-符号集成等未来方向。