IBM Granite team releases LLMON markup language and two preview models
IBM's Granite team published LLMON (LLM Object Notation), a markup language that labels the parts of a model's context — instructions, data, requirements — with reserved special tokens rather than JSON's brackets and quotes. Alongside it, the team released two experimental preview checkpoints derived from Granite 4.1 base models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, plus a GitHub repository containing a toolkit, lossless JSON-to-LLMON converters, templates and workflow examples. Because each instruction is bound by ID to the data it applies to, the models largely disregard instructions hidden in retrieved documents or tool responses; the team reports general benchmark results on par with Granite 4.1 Instruct. Mellea now supports LLMON natively.
Why it matters
Anyone building assistants that read retrieved documents or tool output can mark instructions apart from data instead of flattening everything into one prompt, which the team reports sharply reduces indirect prompt injection success without a conversion pipeline for existing Granite 4.1 3B or 8B users.
Sources
- PrimaryHugging Face
Written by AIToolMaps from the source above and checked against it sentence by sentence.