Blocking in LLMs
Do LLMs know the badness of badity?
Blocking is a well-known linguistic phenomenon. In brief, constructions (including words) are dispreferred if there is an already established construction with the same meaning. For example, childs is dispreferred as a plural of child (even though it follows the regular rule for pluralization in English) because it is “blocked” by the semantically identical word children.
In this project, we investigate the hypothesis that LLMs do not display blocking behavior in the same way the humans do.