What happened
IBM published LLMON, a markup language written for models rather than for humans or old-school parsers. The idea is to stop making the model guess what each piece of its context actually is.
Think about what piles up in a long agent run: a system prompt, earlier conversation turns, instructions about retrieved documents, tool responses, user data, output requirements and reasoning chains. IBM says most of that structure is either implicit or squeezed into formats like JSON, which were never designed for an LLM to read.
LLMON's answer is explicit labeling. IBM describes meaningful tags for semantically coherent components, wrapped in reserved special-token delimiters. Data schemas, nesting included, are spelled out so a causal LLM can read them without ambiguity. IBM also mentions unique IDs, so specific instructions can be pointed at directly.
The format is now integrated into Granite 4.1 models, in 3B and 8B sizes.
Why it matters
IBM's diagnosis is blunt: missing structure is one of the main reasons LLMs, especially smaller ones, are brittle and fail in unexpected ways. Anyone who has watched a small model confidently mangle a tool call knows the feeling.
The security angle is the bigger deal. IBM says the same lack of structure can be exploited by attackers, prompt injection being the example. If the model cannot tell your instructions from text scraped off a random web page, the web page gets a vote.
Reserved delimiters are an attempt to give the model a hard line between "this is the rule" and "this is just data".
Tags and reserved tokens instead of vibes: that is the whole pitch.
Now the caveats. This is IBM's own announcement, and the format lives in two small Granite models. There is no outside validation here and no sign yet that other model families will adopt it. A markup language only becomes a standard when more than one party speaks it.
Still, a vendor attacking prompt injection at the input-format level, rather than with another pile of "please ignore malicious instructions" in the system prompt, is a direction worth watching.
What to do with it
- Try the models.If you build on small open models, test Granite 4.1 3B or 8B on your messiest tool-calling or retrieval flow.
- Audit your context.Map what sits in your prompts today: system rules, tool output, user data. If it all arrives as one blob, you have the problem LLMON targets.
- Treat injection as unsolved.Structure may help, but IBM's claim is mitigation, not immunity, so keep your input filtering and permissions tight.
- Read the IBM post.The Hugging Face write-up has the details on tags, schemas and IDs before you decide whether to build around it.
Ciao, sapiens. When the machines need their own markup language to read their own homework, the least you can do is label your prompts.


