Lingrui Mei, Jiayu Yao, Yuyao Ge, Yiwei Wang, Baolong Bi, Yujun Cai, Jiazhi Liu, Mingyu Li et al.
A survey paper that systematically defines the field of context engineering for optimizing LLM contextual information, and comprehensively covers its components and implementations.
LLM performance heavily depends on provided contextual information, but methodologies for optimization have been studied sporadically, lacking a unified framework. Additionally, there is an asymmetry where models excel at understanding complex contexts but are limited in generating correspondingly sophisticated long-form outputs.
Context engineering is defined as a formal discipline beyond prompt design, categorized into components (context retrieval, generation, processing, management) and system implementations (RAG, memory systems, tool-integrated reasoning, multi-agent systems). Over 1,400 papers are systematically analyzed to present a technical roadmap.
Provides a unified framework for context engineering, identifies a research gap (asymmetry between context understanding and generation capabilities), and suggests future research directions. Offers a valuable reference for both researchers and engineers.