The Magic of LLM Distillation — Rishabh Agarwal, Google DeepMind
Mar 23, 2025 · 46:42
Rishabh Agarwal from Google DeepMind discusses modern LLM distillation, arguing that while simple synthetic data distillation (generating outputs from a teacher and fine-tuning a student) gets 80-90% of the benefit, more advanced on-policy methods (sampling from the student and getting teacher logits) can close the remaining gap by fixing the train-test mismatch. He explains that logit-based distillation is more information-rich but often neglected due to infrastructure complexity, while online distillation is more expensive but crucial for long-horizon or agentic tasks. Agarwal highlights how flipping the KL divergence direction changes behavior (mode-covering vs. mode-seeking) and shows that smaller models can sometimes outperform larger ones in compute-matched distillation due to generating more filtered data. He also connects distillation to speculative decoding, where a distilled student speeds up the teacher, and encourages researchers to question current methods and mine RL literature for ideas.
