This paper systematically compares large-scale language models (LLMs) capable of human conversation to determine whether they experience similar comprehension difficulties to humans. Sentence comprehension data from humans and five state-of-the-art LLMs on seven challenging linguistic structures were collected and analyzed in a unified experimental environment. In particular, LLMs struggle with the sentence "Garden Path," and model performance correlates with human performance in proportion to the number of parameters. Furthermore, performance differences between complex and basic structures are similar for both humans and LLMs, with convergence and divergence observed depending on the model's strength.
