Yannik Keller, Thomas F. Eisenmann
A comprehensive survey reviewing the working principles, emergent capabilities, and ongoing debates about the true understanding of Large Language Models (LLMs).
Despite the advances in Large Language Models (LLMs), fundamental questions about their internal mechanisms, capability limits, and relationship to human cognition remain actively debated. Specifically, the debate continues on whether the capabilities exhibited by LLMs represent genuine cognition or mere pattern learning.
This paper first provides a concise overview of the Transformer architecture and attention mechanism. It then reviews emergent LLM capabilities that resemble human cognition, such as symbolic reasoning, theory of mind, and deception strategies. It also examines explainable AI (XAI) approaches ranging from neuron activation analysis to circuit tracing. Finally, it addresses current debates on genuine understanding in LLMs, advocating for a nuanced discussion that moves beyond simplistic reductionist arguments.
This paper provides a comprehensive synthesis of the current understanding of LLM capabilities and limitations. In particular, it proposes a more sophisticated framework for discussion on LLM cognition that neither ignores the differences between humans and LLMs nor precludes the possibility of AI cognition through overly simplistic arguments. This contributes to setting the research agenda at the intersection of future LLM research and cognitive science.