Researchers propose Graph of Concept Predictors (GCP), a reasoning-aware active distillation framework for Large Language Models (LLMs) that improves performance under limited annotation budgets while yielding more interpretable and controllable training dynamics.