Skip to yearly menu bar Skip to main content


On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes

Rishabh Agarwal · Nino Vieillard · Yongchao Zhou · Piotr Stanczyk · Sabela Ramos Garea · Matthieu Geist · Olivier Bachem

Halle B #76
[ ]
Fri 10 May 7:30 a.m. PDT — 9:30 a.m. PDT


Knowledge distillation (KD) is widely used for compressing a teacher model to reduce its inference cost and memory footprint, by training a smaller student model. However, current KD methods for auto-regressive sequence models suffer from distribution mismatch between output sequences seen during training and those generated by the student during inference. To address this issue, we introduce Generalized Knowledge Distillation (GKD). Instead of solely relying on a fixed set of output sequences, GKD trains the student on its self-generated output sequences by leveraging feedback from the teacher on such sequences. Unlike supervised KD approaches, GKD also offers the flexibility to employ alternative loss functions between the student and teacher, which can be useful when the student lacks the expressivity to mimic the teacher's distribution. Furthermore, GKD facilitates the seamless integration of distillation with RL fine-tuning (RLHF). We demonstrate the efficacy of GKD for distilling auto-regressive T5 language models on summarization, translation, and arithmetic reasoning tasks.

Chat is not available.