Do Large Language Models know who did what to whom?
This addresses a core language understanding problem for AI researchers, providing incremental insights into LLM capabilities.
The study investigated whether large language models (LLMs) can infer thematic roles (who did what to whom) from sentences, finding that while some attention heads capture this information, overall representations reflect syntax more strongly than role assignments, with weaker influence compared to humans.
Large Language Models (LLMs) are commonly criticized for not understanding language. However, many critiques focus on cognitive abilities that, in humans, are distinct from language processing. Here, we instead study a kind of understanding tightly linked to language: inferring who did what to whom (thematic roles) in a sentence. Does the central training objective of LLMs-word prediction-result in sentence representations that capture thematic roles? In two experiments, we characterized sentence representations in four LLMs. In contrast to human similarity judgments, in LLMs the overall representational similarity of sentence pairs reflected syntactic similarity but not whether their agent and patient assignments were identical vs. reversed. Furthermore, we found little evidence that thematic role information was available in any subset of hidden units. However, some attention heads robustly captured thematic roles, independently of syntax. Therefore, LLMs can extract thematic roles but, relative to humans, this information influences their representations more weakly.