AI engineering glossary · LLM training alignment
What is exposure bias?
Exposure bias: A mismatch in models trained with teacher forcing: during training they always see correct previous tokens, but at generation time they see their own possibly wrong outputs, so errors can compound.
Exposure bias: an example
A translation model trained on perfect prefixes drifts off after one wrong word at inference.
Related terms
More LLM training alignment terms
- FLAN
- FLOPs
- Foundation model
- FP8 training
- Full fine-tuning
- Fully sharded data parallel
- Generalised advantage estimation
- Goodhart's law
Where you learn it
Exposure bias is part of the AI Engineering Program, which teaches it with notes, a hands-on lab and an exam. The first lesson is free.