LLMON: An LLM-native Markup Language to Leverage Structure and Semantics at the LLM Interface
The work introduces LLMON, an LLM-native markup language designed to structure text with semantic metadata like instructions or data. This solves the problem of confusing instructions and data in plain text prompts, which can lead to security risks. LLMON improves model training by providing explicit structural signals and enhances inference by allowing systems to isolate specific instructions for reliable execution.