Back to search

University of Minnesota

Cognitive Perspectives on Intelligence in Large Language Models

Abstract

dc:description.abstract

Large language models (LLMs) are the backbone of recent consumer-facing artificial intelligence systems. Their linguistic intelligence encodes not only language features like syntax and semantics but also world knowledge, making LLMs usable as conversation partners and assistants during activities like writing and programming. However, the extent to which LLMs learn other forms of intelligence – like problem solving or creativity – remains unclear. In some tasks, models can produce superhuman results, whereas in others their responses are lacking or erroneous. Because precise LLM capabilities are unclear and poorly understood, it is difficult to predict when they will succeed or fail. What is the nature of cognition and language understanding in LLMs, and how does it differ from that of humans? This thesis takes an interdisciplinary approach to answering these questions, drawing on methodology from the cognitive sciences. First, we investigate linguistic competencies in LLMs. We find that LLMs diverge from humans in several pragmatic discourse interpretation tasks, despite their high degree of functional linguistic ability. Second, we investigate non-linguistic competencies in LLMs, finding that LLMs have deficiencies in performing laboratory tasks designed to test goal-oriented cognitive functions in humans, and we find procedural differences in how LLMs and human experts approach complex knowledge work in law and the sciences. To support this work, we also present novel data collection methods that capture cognitively informed signals alongside human data. By evaluating cognitive competencies of LLMs across several domains, this work lays a foundation for developing a theoretical framework of cognition in LLMs and AI systems more broadly.

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • de Langis, Karin

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Citation
de Langis, Karin. (2026). Cognitive Perspectives on Intelligence in Large Language Models. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/283464.

Chain of custody

source
Harvested from
University of Minnesota
Base URL
conservancy.umn.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

de Langis, Karin. Cognitive Perspectives on Intelligence in Large Language Models. 2026. https://hdl.handle.net/11299/283464