Does GPT-5.2 Understand Discourse Coherence?Evidence from the Referential Structure Effect

초록

The present research investigates discourse competence of GPT-5.2 using the referential structure effect, which is a coherence-based discourse constraint on pronoun resolution grounded in Centering Theory. In human comprehension, Two-Pronoun structures (e.g., …he verbed him) tend to increase subject-antecedent interpretations of the subject pronoun relative to One-Pronoun structures (e.g., …he verbed Joe), because object interpretations in Two-Pronoun structures can induce a coherence- disrupting promotion/demotion configuration. Thus, the present work aims to examine (i) whether GPT-5.2 exhibits this referential structure effect and (ii) whether it replicates the implicit causality (IC) bias, which involves causal inferences in subsequent discourse. Adopting Song and Kaiser’s (2023) work with human participants, this study manipulated IC verb type (IC1 vs. IC2) and referential structure type (One-Pronoun vs. Two-Pronoun). To control for potential effects of prompt format, three prompting formats (A/B, Repeat, Compare) were adopted with option-order counterbalancing. Across formats, GPT-5.2 consistently replicated IC-bias effects found in previous work on humans and LLMs. However, the referential structure effect showed human-like patterns only partially and varied depending on the prompt format and verb type. These results suggest that GPT-5.2 is not consistently sensitive to subtle discourse coherence constraints and that prompt format can modulate LLMs’ discourse competence.

키워드