Proceedings of the Annual Meeting of the Cognitive Science Society

We examined whether large language models (LLMs) responded to syntactic prominence of discourse entities and pragmatic biases in Turkish subject anaphora resolution similar to human judgments. Using an offline comprehension task with native speakers, we showed that null and overt pronouns responds to syntactic prominence and pragmatic bias. We then evaluated several autoregressive LLMs on the same materials. While GPT-4o correlated with human responses, LLaMA-4 more closely approximated the interaction between syntactic prominence and pragmatic biases observed in human data, raising questions about the relationship between model performance, scale, and human fit. We also found that all models mostly differed from humans in response variability, with model responses tending to be more deterministic. Finally, the results indicated partial but limited alignment between human and model anaphora resolution in Turkish.