With the rapid development of large language models and agent technologies, artificial intelligence is evolving from an information-processing tool into intelligent systems capable of autonomous planning, decision-making, and complex task execution. These systems have shown broad application potential in scientific discovery, software development, data analytics, recommender systems, embodied intelligence, and beyond. However, current mainstream approaches still rely heavily on massive datasets, large-scale models, and continuously increasing computational resources. As the costs of model training, inference, and deployment continue to rise, enabling efficient learning with limited data, supervision, and interaction resources has become a key challenge for the development of next-generation agent systems.

Recently, the AI research team at the Beijing Institute of Mathematical Sciences and Applications (BIMSA) completed a survey paper titled Beyond Scaling: A Survey of Data-Efficient Learning for LLM Agents. The paper systematically reviews the emerging research direction of Data-Efficient Agentic Learning and constructs a unified research framework from three dimensions: Experience Augmentation, Agent Structural Design, and Learning Paradigms. The paper has been accepted by the International Joint Conference on Artificial Intelligence (IJCAI 2026), a leading international conference in artificial intelligence.

The authors of the paper include Yaqing Wang, Associate Professor at BIMSA; Zhenlin Luo, Ph.D. student at BIMSA and the School of Statistics and Big Data, Renmin University of China; Peiyao Zhao, Postdoctoral Researcher at BIMSA and the Yau Mathematical Sciences Center, Tsinghua University; Yunfeng Cai, Professor at BIMSA; and Quanming Yao, Associate Professor at Tsinghua University.

Addressing the data, feedback, and interaction costs faced by large language model agents in real-world scenarios, the paper reexamines the development path of agent learning from the perspective of Data Efficiency. It brings together research efforts that have previously been scattered across areas such as experience reuse, memory augmentation, structured context, test-time learning, preference learning, and agent architecture design, and unifies them under the framework of Data-Efficient Agentic Learning. Compared with the scaling paradigm, which primarily relies on increasing model size, this framework focuses more on how agents can continuously improve their capabilities by leveraging existing experience, limited feedback, and structured knowledge. In doing so, it aims to reduce learning costs, improve generalization, and enhance agents’ adaptability and long-term evolution in real-world environments.

The paper further provides a systematic review of recent progress from the three dimensions of experience augmentation, agent structural design, and learning paradigms. In terms of experience augmentation, it discusses techniques such as experience retrieval, case reuse, long-term memory, and simulated interaction. In terms of agent structural design, it analyzes representative frameworks including modular agents, multi-agent collaboration, tool-augmented agents, and external knowledge integration. In terms of learning paradigms, it reviews the current development and challenges of in-context learning, test-time adaptation, preference learning, and budget-constrained reinforcement learning. Through this framework, the paper provides a systematic perspective for understanding how large language model agents learn, adapt, and evolve under limited-resource conditions.

The research team believes that as large language model agents move from laboratory settings toward industrial applications, data efficiency will become a core metric as important as model capability itself. Data-Efficient Agentic Learning not only provides a unified framework for the academic community to understand the mechanisms of agent learning, but also offers practical guidance for the design, deployment, and continuous optimization of enterprise-level agent systems. For example, in scenarios such as enterprise knowledge assistants, software engineering agents, GUI agents, scientific AI, and personalized recommendation systems, it is often difficult for organizations to continuously obtain large-scale, high-quality labeled data. Instead, such systems need to improve their capabilities through historical experience, user feedback, structured knowledge, and limited interactions. The Data-Efficient Agentic Learning framework provides a systematic technical roadmap for these real-world production scenarios and is of significant value for reducing deployment costs, improving iteration efficiency, and strengthening long-term service capabilities.

To further advance Data-Efficient Agentic Learning and promote academic exchange and collaboration among researchers in related fields, Associate Professor Yaqing Wang and Assistant Professor Nan Yin from The Education University of Hong Kong are jointly organizing the ADMA 2026 Special Session on “Data-Efficient Agentic Learning for Data Mining” (DEAL-DM). The special session focuses on the intersection of data-efficient learning and agent technologies, with particular attention to frontier topics such as experience reuse, structured context, memory augmentation, personalized agents, test-time learning, budget-efficient reinforcement learning, and scientific AI. It aims to explore how to build more efficient, reliable, and adaptive agent systems under limited data, feedback, and interaction resources.

DEAL-DM seeks to bring together researchers from data mining, machine learning, large language models, agent systems, recommender systems, scientific AI, and related application areas to jointly explore the theoretical foundations, key technologies, and practical applications of Data-Efficient Agentic Learning, and to promote the transition of this emerging direction from academic research to industrial deployment. Accepted papers will be included in the ADMA 2026 conference proceedings and published in Springer’s Lecture Notes in Computer Science (LNCS) series. The special session is now open for submissions worldwide, with a paper submission deadline of June 26, 2026. The research team warmly welcomes researchers from China and abroad to submit their latest work, exchange new findings, and jointly contribute to the development of this rapidly growing research direction.