Technology releaseNovember 17, 2026Редакция 105.kz
LLMON: Markup Language for LLMs in Granite 4.1 — Less Prompt Injection, Structured Output
IBM has released experimental Granite 4.1 3B and 8B models with LLMON, a markup language that explicitly separates instructions from data. This reduces the risk of indirect prompt injection and simplifies structured output for industrial LLM agents.
Key takeaway. For an engineer in Kazakhstan: LLMON is an experimental approach to structuring LLM context, where instructions, data, and requirements are encoded in separate blocks. Granite 4.1 LLMON-preview models show comparable general capabilities and a strong reduction in indirect prompt injection success on the OPI benchmark, but these are previews, not for production use. A pilot is possible in a closed loop after validation on Russian-language data.
What happened
On November 12, 2026, a material about LLMON (LLM Object Notation) was published on the Hugging Face blog. LLMON is a markup language oriented to causal language models. Unlike JSON, where the structure is set by symbols {, \", :, LLMON uses reserved special tokens. This allows separating structural delimiters from content and explicitly marking block boundaries. The authors released two experimental preview models based on Granite 4.1: granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview. The blog states that on the OPI benchmark, LLMON models avoid 47% more indirect prompt injection attacks than the best frontier model tested, and 95% more than the best open-source model. Overall performance is comparable to Granite 4.1 Instruct. The models, JSON-LLMON converters, examples, and tools are available in the GitHub repository; there is also support in Mellea.
What this means for enterprises in Kazakhstan
For plants and production sites where corporate LLMs run in a closed loop, the key value of LLMON is explicit separation of instructions, data, and requirements. In industrial scenarios, the model often receives documents from different sources: work permits, repair records, inspection reports, contractor letters, data extracts from the process control system (APCS). An indirect injection can be hidden in such text; LLMON reduces the probability that the model interprets data as a command. This is important for HSE reporting and maintenance regulations, where the cost of error is high.
The second aspect is structured output. LLMON allows defining typed blocks, for example, event, equipment unit, recommendation, and linking them to specific instructions. Such output is easier to integrate with MES, ERP, or a shift log than free text or JSON, which the model may generate with grammar violations. For Kazakhstan enterprises working according to ISO and GOST standards, this can simplify automatic classification of defects, extraction of parameters from reports, and generation of brief summaries for the dispatcher.
However, the source material does not provide quality assessments for Russian and Kazakh. For use in Kazakhstan, you will need your own validation on domain corpora, including technical documentation and professional jargon (assessment). The 3B and 8B models are relatively compact, allowing on-premise deployment without extreme computational costs (assessment).
How to apply: 8–12 week pilot
- Audit text flows. Select 2–3 processes where an LLM could take over structured data extraction: processing work permits, classifying repair requests, shift summaries. Assess which instructions and data are currently mixed in a single text.
- Define the target output fields. For example, for a defect report: equipment code, defect type, criticality, suggested action. This will become the schema for LLMON markup.
- Deploy the preview model in a test environment. Use a closed server, with no data leaving the perimeter. Test on anonymized historical documents to see how the model handles Russian terminology. If necessary, prepare a domain dataset for a LoRA adapter.
- Measure the effect. Compare the share of correctly filled fields and processing time with the current manual process or with JSON prompting. Separately test robustness against intentionally injected instructions in test documents.
- Make a decision. If quality is acceptable, proceed to a pilot in a limited environment under NDA. For more on implementation approaches, see the Corporate LLMs page and the Industrial enterprises sector section.
Limitations and risks
LLMON and the Granite 4.1 LLMON-preview models are experimental and not intended for production use. The authors explicitly state: ready for exploration and evaluation, but not for production. Protection against indirect prompt injection has been demonstrated on static benchmarks OPI, ASB, and StruQ; work is still ongoing against adaptive attacks that unfold over many dialogue turns. LLMON behavior is learned, not enforced; failures are possible.
For Kazakhstan enterprises, the main risk is the lack of data on quality in Russian and Kazakh, as well as on highly specialized technical vocabulary. Small models may be inferior to larger ones in complex reasoning, although overall performance is claimed to be on par with Granite 4.1 Instruct. Integration with existing JSON pipelines is possible through converters, but losses during reverse conversion need to be checked separately. Regulatory requirements for data localization and information security of national companies mean that all experiments must take place in a closed loop; cloud services for production data are unacceptable.
Sources
- LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B · 2026-11-12T19:07:35.000Z
LLMON, on the other hand, is an XML-like markup whose structural delimiters are reserved special tokens rather than characters, so structure and content use separate parts of the vocabulary.
- Иллюстрация: сгенерирована 105.kz (z-image, Apache 2.0)
This material was prepared with AI from the sources listed above. Spotted an inaccuracy? Write to info@105.kz.