Tag
This paper investigates how transformer language models learn 'impossible' languages with unnatural properties, finding that while grammatical sensitivity degrades gradually, generative production shows pronounced failures, suggesting a linking hypothesis for non-attestation.