On October 24, 2026, a research paper titled “Once a Tree, Always a Tree? Cross-lingual Transfer of Hierarchical Generalization in Language Models” was officially published on the arXiv platform under the identifier 2610.30960v1. The study focuses on a core mechanism that remains unclear in cross-lingual transfer within language models (LMs): whether models can effectively transfer evidence of hierarchical grammatical structures learned in the source language (L1) to the target language (L2), thereby supporting L2’s capacity for hierarchical generalization.
To test this hypothesis, the research team devised a novel “Cross-Language Poor Stimulus” (CPOS) experimental paradigm: they deliberately introduced ambiguity in both linear and hierarchical grammatical rules within an L2 corpus, while rigorously controlling the amount of positive evidence in L1 that supports hierarchical rules. The results show that language models can partially—but not fully—transfer L1’s hierarchical‑structure evidence to L2, yielding only a weak yet detectable hierarchical preference; moreover, the strength of this transfer varies across language pairs.
Further ablation analyses revealed that lexical overlap is a critical condition for cross-linguistic, hierarchical syntactic generalization and transfer—this finding not only explains the origins of language‑pair dependencies observed in the main experiment but also demonstrates that, during continued training on new languages, language models do not automatically transfer linguistic evidence; rather, such transfer is highly contingent on specific conditions and exhibits significant limitations.