About the journal
Browse
Collections
Multimedia collections
Authors & reviewers
Training Small Language Agents: A Survey of Methods and Opportunities
Chengyu WANG , Taolin ZHANG , Richang HONG
Large language model (LLM) agents have demonstrated strong capabilities in reasoning, planning, tool use, and interactive decision-making. However, their training and deployment often depend on substantial support in model scale, context length, inference budget, latency, memory, and external tools. This has motivated growing interest in small language agents, which we define not simply as agents built on small language models, but as agentic systems trained to deliver useful multi-step behavior under explicit deployment constraints. Despite rapid progress, research in this area remains scattered across adjacent topics such as small language models, agent fine-tuning, knowledge distillation, process supervision, and reinforcement learning, and a dedicated survey is still lacking. In this survey, we present a structured review of training small language agents as a distinct learning problem shaped by limited model capacity, bounded support, and deployability requirements. We first clarify the foundations of small language agents, discuss their defining characteristics, and distinguish them from small language models, general LLM-based agents, and efficient agentic systems more broadly. We then organize existing methods according to the main stages and mechanisms of training, covering supervised fine-tuning as the primary pre-deployment foundation; trajectory learning and process supervision as finer-grained forms of supervised agent training; knowledge distillation as a cross-cutting mechanism for transferring agentic capabilities; reinforcement learning as interaction-based policy refinement; and continual improvement as a post-deployment lifecycle stage. Throughout, we compare representative systems in terms of backbone scale, target capabilities, supervision sources, deployment constraints, and reported results, and derive practical guidelines for selecting among training approaches under different deployment budgets. We further review the resources and evaluation protocols most relevant to this setting, including human demonstrations, teacher-generated data, synthetic tasks, simulated environments, interaction trajectories, and deployment-aware assessments of efficiency and robustness. Finally, we summarize the main challenges and highlight promising directions for building effective, reliable, and deployable small language agents.
small language agents / small language models / agent training / knowledge distillation / reinforcement learning
Higher Education Press 2026
/
| 〈 |
|
〉 |