Event

Central South University and Other Institutions Release a System-Level Agent Distillation Survey

1 reports · 1 sources

A research team from institutions including Central South University, the Beijing Academy of Artificial Intelligence, The Hong Kong Polytechnic University, and Beijing University of Posts and Telecommunications jointly released a 57-page survey on system-level agent distillation, titled Distilling Agentic Systems: A Roadmap across Models, Artifacts, and Harnesses, available at https://arxiv.org/abs/2609.36630. Taking complete agents as both teacher and student, the survey establishes a unified definition, taxonomy, and evaluation framework around how task-solving knowledge is distilled, retained, and used for subsequent tasks, and further discusses the transfer and composition of knowledge across different carriers. Using where knowledge is retained in the student as the basis for classification, the survey organizes related methods into three main lines: Models, reusable Artifacts, and execution control Harnesses. The survey also builds an evaluation framework connecting knowledge retention, causal contribution, and actual task benefits, systematically examining what the student retains, how these changes affect execution, and whether the related benefits continue to hold in new tasks and deployment environments.