IBM Releases LLMON Markup Language in Granite 4.1 3B and 8B Preview Models

At a glance

IBM's Granite team released LLMON, a markup language that helps large language models better recognize the structure of their context. They also released two experimental preview models built on it.

Key points

1. 1

   LLMON tags instructions, data, and requirements so each is clearly marked.

2. 2

   This separation makes the models mostly ignore hidden malicious instructions in user data.

3. 3

   In some tests, attack success rates dropped to nearly zero.

4. 4

   A converter also turns JSON into LLMON without loss.

Why it matters

It can reduce security risks and malfunctions when AI apps handle outside documents or user input.