The paper behind LLMON starts from a simple observation: textual LLMs offer a familiar interface, but the information we feed them is rarely just a flat string. It often carries hierarchy, relationships, and semantic meaning that plain text can only approximate. The authors propose a markup language designed specifically for LLMs, one that encodes such structure directly in the input and output streams.
By making structure explicit, LLMON could help models better understand complex instructions or data without requiring a separate parser or a non-textual interface. The approach is aimed at the point where humans and software meet the model, rather than at the model's internal architecture.
Because the abstract is cut off, the article stops short of describing how LLMON is implemented or whether it has been tested. Still, the motivation is clear: if text is the universal interface, giving it a native way to carry structure may improve both reliability and expressiveness.