Do LMs have word order preferences & are they architectural or data-driven?
In our paper, the same architecture robustly prefers left-branching artificial languages, but on natural languages, as training data grows, right-branching SVOs win 🧵
Accepted to #EMNLP2026 main 🎉
