LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

Practical AI: Tools, Models & Frameworksllm

What changed

IBM released LLMON (LLM Object Notation), an XML-like markup language using reserved special tokens instead of characters to explicitly structure LLM context. Two experimental Granite 4.1 preview models (3B and 8B) are now trained to consume and generate LLMON, along with a GitHub toolkit for conversion and workflow templates.

Why it matters

LLMON enables robust instruction/data separation that reduces indirect prompt injection attack success rates to near-zero on benchmarks while maintaining general capability parity with standard instruct models. Builders can bind instructions to specific data payloads, ignore unrelated context blocks, and define domain-specific schemas with LoRA adapters—eliminating fragile JSON parsing and implicit context structure.

How it compares

Unlike standard chat templates that flatten all components into one string, LLMON enriches chat templates with explicit, addressable blocks and execution bindings. Unlike JSON (which was designed for parsers), LLMON's structural delimiters are single tokenizer tokens while content uses separate vocabulary, making structure and semantics directly learnable by causal models.

Sources