- What changed
- Researchers causally isolated a rare Japanese inflection subclass accounting for under 1% of training data but generating up to 43% of residual errors in character-level Transformers.
- Why you should care
- Aggregate accuracy metrics can hide critical failures caused by interactions between orthographic representations and rare morphological inflection patterns.
- Your move
- Watch. High benchmark scores can obscure severe localized error concentrations.
- What to watch next
- Independent replication across non-Japanese orthographies or larger pretrained models evaluating whether rare morphological failure modes persist at scale.
- Event
- research
- Event date
- Sep 21, 2026
- Relevant to
- General AI readers
