Interpretable Humans, Alien LLMs: Expert Analysis of Latent Structures in Assessment Responses

Evaluating large language models (LLMs) often relies on human-designed tests, assuming that AI agents and human learners employ similar cognitive constructs to arrive at correct answers. This paper challenges that assumption by investigating whether the latent statistical factors governing LLM performance carry the same substantive, human-interpretable meaning as the cognitive constructs governing human learners. The authors administered established chemistry and quantitative reasoning assessments to both human students and six multimodal LLMs. They then applied Exploratory Factor Analysis (EFA) separately to each group's response data to uncover underlying factor structures.

To determine whether these statistical factors correspond to meaningful pedagogical or cognitive concepts, the researchers used a blind evaluation framework. Subject-Matter Experts (SMEs) were given the factor loading graphs without knowing whether the factors originated from human or LLM responses. The experts examined the item content associated with each factor, looking for shared epistemic, cognitive, or psychometric characteristics. While SMEs successfully identified clear pedagogical and problem-solving strategies for all human-derived factors, they faced a stark interpretive barrier when evaluating the LLM data.

The results revealed a profound gap between human and machine response patterns. For quantitative reasoning, SMEs could not ascribe meaningful interpretations to any of the LLM-derived factors, despite easily identifying strategies like data-graph reading and test-taking heuristics in human factors. In the chemistry assessment, experts could interpret only two out of the four LLM factors. This indicates that while LLMs can achieve high aggregate accuracy on human exams, their internal response patterns frequently operate on statistically opaque mechanisms that diverge fundamentally from human reasoning.