Stack-Based LMs Struggle with Cross-Serial Dependencies in Artificial Languages
October 2, 2026
Stack-based language models show improved generalization on hierarchical patterns when using limited working memory, yet they still fail to master cross-serial dependencies. This finding suggests that structural complexity limits in current architectures may influence the acquisition of typological commonalities seen in natural languages.
HOW THIS AFFECTS YOU
●
researcherYou can use these findings to investigate how working memory constraints influence inductive biases in hierarchical modeling.