Dev News Daily ENDE

IBM releases LLMON, a markup built from reserved tokens, with two Granite 4.1 preview models

IBM's Granite team published LLMON on 12 November, short for LLM Object Notation and pronounced lemon. It is a markup language meant to be read by language models rather than by parsers, and it arrives with two experimental models derived from Granite 4.1 base checkpoints, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, plus a GitHub repository with converters, templates and example workflows.

The problem the authors describe is familiar to anyone building agents. Over a long session the context fills with a system prompt, earlier turns, retrieved documents, tool output, output constraints and reasoning. Most of that structure is either implicit or written in JSON, and the post argues that JSON was designed for parsers. A tokenizer does not respect its grammar, the same quote character opens text and closes a field, and nesting has to be inferred by counting brackets over long spans, which transformers handle poorly.

LLMON looks a little like XML, but its delimiters are reserved special tokens rather than ordinary characters, so markup and content draw on different parts of the vocabulary. Each span states its tag at the start, nested structure is declared locally, and every component gets a unique ID. An exec block at the end of the rendered prompt then says which instruction to run, on which data and under which requirements, by referring to those IDs. In the chat template a message part can be flagged exec True or False; parts set to False still appear in the context but nothing points at them.

IBM's headline claim concerns indirect prompt injection. Because the model is trained to act only on spans bound into the exec block, instructions hidden inside user data are mostly ignored. On some benchmarks the post reports attack success close to zero for the LLMON models, while ordinary instruction-tuned models are hijacked with high probability, and it says general capability stays in line with Granite 4.1 Instruct. The team also says unrelated passages and old turns are less likely to leak into answers, since each instruction is tied to the data it applies to.

LLMON does not replace chat templates, the authors stress, it adds meaning inside them: roles say who is speaking, LLMON says what each part is. Teams with JSON pipelines can convert in both directions without loss, and the Mellea library already wraps its model calls in LLMON automatically.

IBM releases LLMON, a markup built from reserved tokens, with two Granite 4.1 preview models
IBM releases LLMON, a markup built from reserved tokens, with two Granite 4.1 preview models — Dev News Daily

Why it matters

Prompt injection is usually handled outside the model, with filters and wrappers. LLMON moves part of the separation into training and tokenization. The models are previews and the strongest numbers are reported for selected benchmarks, so the claim needs independent testing before anyone treats the format as a defence.