Scaling LLM Agent Learning with Data Synthesis: A Comprehensive Survey

Hanrong Zhang, Yankai Chen, Shicheng Fan, Dehai Min, Shaowen Chen, Huanhuan Ma, Zhaofen Wu, Jie Yang, Bowei He, Jikun Kang, Kening Zheng, Xi Chen, Chunyu Miao, Fulin Lin, Wei-Chieh Huang, Jiayu Zhou, Haolun Wu, Liancheng Fang, Hong Kang, Langzhou He et al. (11 additional authors not shown)

Abstract:

LLM agents are transforming language models from chatbots into interactive systems that operate through memory, tool use, and external environments. Training such agents requires coherent goals, multi-step trajectories, feedback, and outcomes derived from long-horizon interactions, yet these data are difficult to collect at scale because environments are costly, human oversight is limited, and execution traces are complex.
Therefore, data synthesis offers a practical route to scaling agent learning by sampling supervision from source distributions such as teacher models, seed-conditioned generators, simulators, judges, or executable environments. Despite growing interest, no survey has systematically organized synthetic data for LLM agents.
To bridge this gap, we present a comprehensive survey that classifies synthesis methods for LLM agents, by their outputs within the agent--environment loop, including task specifications, agent trajectories, feedback signals, and environments.
We examine the lifecycle of these artifacts, from grounding and quality control to their utilization in learning, evaluation, and downstream applications, and identify key challenges for reliable and scalable agent-data synthesis.