This paper shows that LLMs do not really understand language, even though popular opinion suggests that they do. Specifically, it highlights that LMMs are statistical models and that the limit of 'understanding' language is of learning how words relate to each other, not why the relate to each other.

This is an early and incomplete paper which seems to have been completed subsequently as a separate paper called "Stochastic LLMs do not Understand Language: Towards Symbolic, Explainable and Ontologically Based LLMs" which is also attached.

Saba, W.S. (2024) ‘LLMs’ Understanding of Natural Language Revealed’. arXiv. Available at: https://doi.org/10.48550/arXiv.2407.19630.

Saba, W.S. (2023) ‘Stochastic LLMs do not Understand Language: Towards Symbolic, Explainable and Ontologically Based LLMs’. arXiv. Available at: https://doi.org/10.48550/arXiv.2309.05918.

This is a paper that discusses how RAG is an antidote to the static offline nature of LMMs and how they can be used to provide context and dynamic content, which can also combat hallucination (reliance only on training data).

This is co-written by the same researcher (Walid S. Saba) who wrote:

1. LLMs' Understanding of Natural Language Revealed
2. Stochastic LLMs do not Understand Language: Towards Symbolic, Explainable and Ontologically Based LLMs

Reference:

Church, K.W. et al. (2024) ‘Emerging trends: a gentle introduction to RAG’, Natural Language Engineering, 30(4), pp. 870–881. Available at: https://doi.org/10.1017/S1351324924000044.