The Anatomy of a Chinese Knowledge Distillation Campaign

Abstract

Chinese AI developers may not need to develop agentic capabilities on their own. “Black-box knowledge distillation” — a technique where a weaker “student” model learns by querying a more capable “teacher” model — has emerged as a technique to “copy” the capabilities from closed-weights models. To understand the role of distillation in model development, we embarked on a literature review of academic research and recent reports on how Chinese model developers engage with the technique. We will present how Chinese labs are believed to circumvent guardrails, location policies, and account restrictions to generate synthetic data that improves their models.

We break the distillation process itself into five critical steps: select, seed, steer, generate, and train. By understanding these components of distillation, we can begin to understand how to detect and respond to attempts to copy frontier capabilities. Furthermore, we discuss the strategic implications for AI security, highlighting the difficulties developers face in distinguishing adversarial distillation campaigns from legitimate industrial usage.