Why Knowledge Distillation Works in Generative Models: A Minimal Working Explanation
This work provides a foundational explanation for a core technique in generative AI, addressing a gap in theoretical understanding for researchers and practitioners.
The paper tackled the problem of understanding why knowledge distillation improves generative models by analyzing its mechanisms, demonstrating through simulations and large-scale language models that it induces a precision-recall trade-off, where student models gain sample quality at the expense of coverage, with empirical validation showing benefits in scenarios prioritizing quality over diversity.
Knowledge distillation (KD) is a core component in the training and deployment of modern generative models, particularly large language models (LLMs). While its empirical benefits are well documented -- enabling smaller student models to emulate the performance of much larger teachers -- the underlying mechanisms by which KD improves generative quality remain poorly understood. In this work, we present a minimal working explanation of KD in generative modeling. Using a controlled simulation with mixtures of Gaussians, we demonstrate that distillation induces a trade-off between precision and recall in the student model. As the teacher distribution becomes more selective, the student concentrates more probability mass on high-likelihood regions at the expense of coverage -- a behavior modulated by a single entropy-controlling parameter. We then validate this effect in a large-scale language modeling setup using the SmolLM2 family of models. Empirical results reveal the same precision-recall dynamics observed in simulation, where precision corresponds to sample quality and recall to distributional coverage. This precision-recall trade-off in LLMs is found to be especially beneficial in scenarios where sample quality is more important than diversity, such as instruction tuning or downstream generation. Our analysis provides a simple and general explanation for the effectiveness of KD in generative modeling.