Agentic Environment Engineering for Large Language Models: A Survey of Environment Modeling, Synthesis, Evaluation, and Application

The paper presents a comprehensive survey on agentic environments for large language models (LLMs), detailing the environment engineering lifecycle, which includes modeling, synthesis, evaluation, and application. It categorizes environments based on eight attributes and domains, introduces two paradigms for automated environment synthesis (symbolic and neural), and discusses evaluation methods and applications linked to agent-environment co-evolution. This work is significant for practitioners as it identifies pathways for agent evolution and suggests future directions like Environment-as-a-Service and Multi-agent Environments, which could enhance the adaptability and performance of LLMs in dynamic settings.