LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B
Web page · https://huggingface.co/blog/ibm-granite/llmon-an-llm-native-markup-language-now-in-granite
First seen by Oriel on .
AI summary
This post introduces LLMON, an LLM-native markup language and two Granite 4.1 preview models that explicitly structure context to improve prompt-injection resistance without reducing general capability.
- LLMON uses reserved special-token delimiters, typed tags, local nesting, and IDs linking instructions to exact data, requirements, and output blocks.
- On OPI, the previews avoided 47% more attacks than the best frontier model and 95% more than the best open-source model; adaptive attacks remain under development.
- Granite 4.1 3B and 8B LLMON models perform comparably to Granite 4.1 Instruct across standard benchmarks and work through their supplied chat template.
- Training used serialized instruction data, decoy blocks, five GRPO stages, and a final DPO stage; the resulting privilege behavior is learned, not enforced.
- JSON can be converted losslessly, while custom schemas support LoRA adapters demonstrated on ScienceWorld; the release is experimental and not for production.
#llmon #prompt injection #granite #llm security
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Blog | https://huggingface.co/blog/ibm-granite/llmon-an-llm-native-markup-language-now-in-granite | https://huggingface.co/blog/ibm-granite/llmon-an-llm-native-markup-language-now-in-granite Read discussion |