Distillation Learns Hesitation, Not Answers: Why Small Models Get Strong and Then Collapse

大模型之路 · wechat · 2026-08-27

A systematic guide to knowledge distillation: the student learns the teacher's full probability distribution ("85% cat, 10% lynx") — that 10% encodes inter-class similarity, Hinton's dark knowledge, amplified via temperature scaling (typically 2-20).

Orthogonal to quantization/pruning: quantization lowers precision and pruning removes redundancy within the same model, while distillation trains a new model; the right order is distill first, then quantize (DistilBERT: 40% smaller, 60% faster, retains 97% of BERT's GLUE score).

Why distilled models collapse on long tasks: offline distillation suffers exposure bias — the student only saw the teacher's polished outputs, so errors compound with O(T²) growth on multi-step generation. On-policy distillation has the student generate first while the teacher gives per-token distributions, reducing error to O(T) with far fewer training rounds than RL. A survey (arXiv:2604.00626) organizes these methods via an f-divergence framework across feedback signal, teacher visibility, and loss granularity.

Two hard constraints: the teacher is the ceiling, and distillation transfers domain-specific response behavior, not the teacher's full knowledge — its errors are inherited and amplified. Per DeepSeek-R1's report, large-scale RL on small models may underperform distillation: RL grows new capabilities in frontier models; distillation cheaply copies existing ones.

Practical path: log real production queries answered by the teacher as distillation data → fine-tune an open-source student → evaluate multi-turn and long responses for degradation → then graduate to on-policy distillation.

Original post →

More from Models

Models channel →