A new arXiv paper wades into the debate over what large language models actually do with language. The paper contends that the current generation of LLMs does not understand language in any meaningful sense, but rather "does" it. That distinction is central to the argument.
The paper pushes back against the common view that LLMs form some kind of distributed representation of symbols. Instead, the claim is that LLMs are capturing something else—though the available abstract cuts off before spelling out exactly what that alternative is.
If the argument holds, it would shift how researchers talk about model internals: less about semantic understanding, more about behavioural mastery. The paper is posted on arXiv under the title "Communication, Vocalisation, and Language."