Action nominal constructions (ANCs) are a subtle area of grammar: they turn events into nouns while borrowing argument-marking patterns from either clauses or ordinary noun phrases. A new preprint on arXiv, "From Parallels to Divergence: Language Models' Learning of Action Nominal Constructions," sets out to probe how well language models acquire these constructions.
The abstract indicates the work focuses on how different grammatical configurations of ANCs are learned, though the full details of the experiments are not included in the summary. The title hints that models may initially treat these configurations in parallel ways before diverging in their behavior, but the abstract itself does not state specific results.
As the only source, this preprint offers a limited but clear research direction: understanding where language models succeed or fail at grammar that sits between clausal and nominal syntax. Further reading of the full paper would be needed to assess the claimed divergence and its implications for model evaluation.