research
Knowledge That Is Linear Before Fine-Tuning Generalizes After It
Researchers distinguish active facts that support generalization from inert facts that can be recalled but do not. They find that how linearly readable an original relation is in pretrained representations strongly predicts whether a new relabeled relation generalizes to held-out subjects.
- Why it matters
- It suggests that generalization depends less on fact recall than on how readily shared structure can be read out linearly.
- Who is affected
- Language model researchers are affected.