Injecting 1% targeted synthetic data into GPT-2's pre-training substantially improves performance on 8 of 9 failing BLiMP grammatical paradigms, indicating data scarcity causes formal linguistic failures.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
No citing papers match the current filters.