Is our understanding of LLMs too narrow?

Large language models (LLMs) are trained through statistical prediction of token sequences. Whether token prediction alone adequately describes how mature models behave remains unclear. We examined this question by asking eight commercially available LLMs to take turns continuing stories. This was done under two conditions: (1) heavily corrupted spelling and writing patterns and (2) an invented pseudo-language. Neither condition allowed simple continuation of familiar language patterns. These tasks were designed to test whether the LLMs would preserve meaningful and coherent language. Across both tasks, all eight models extended corrupted spelling to new words, maintained the emerging language systems, generated novel vocabulary and concepts, and continued building coherent narratives without explicit grammar instruction or correction. These findings suggest that simple token prediction may be insufficient as a description of what mature models do. The observed behavior is more easily explained by internal representations that reconstruct and stabilize meaning than by simple continuation of familiar language patterns. In simple words: LLMs may develop their own form of understanding during training. Current descriptions of LLM capability may therefore be unnecessarily narrow.