HyperAIHyperAI

Command Palette

Search for a command to run...

Granite 4.1 3B, 8B Preview Models Add LLM-Native LLMON Markup

IBM has introduced LLMON, or LLM Object Notation, an LLM-native markup language designed to make the structure and meaning of model inputs explicit. The company also released two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, along with a GitHub repository containing a toolkit, converters, examples, and workflow templates. LLMON targets a recurring weakness in causal language models: long contexts contain system prompts, retrieved documents, tool responses, user data, constraints, and reasoning chains, but their roles are often implicit or encoded in formats such as JSON. To a model, ordinary structural characters can look identical to content characters, making instruction, data, and requirements harder to distinguish. LLMON instead uses tagged markup wrapped in reserved special token delimiters, so structure and content occupy separate vocabulary items. Tags delimit blocks, schemas can describe nesting, and unique block identifiers allow instructions to point to specific data and requirements. A generated exec block then ties an instruction, its inputs, its constraints, and the output slot together. IBM said LLMON is intended to complement, not replace, chat templates. The preview models ship with a template that renders user messages into named blocks and an execution section. If a turn labels components as instruction, data, and requirement, the template marks which items should be executed and which are only background. The approach is especially relevant to indirect prompt injection, where hidden commands are inserted into retrieved documents, tool results, or user-provided data. By giving the model a clearer separation between authoritative instructions and data to be read, LLMON can reduce the chance that it will carry out malicious embedded directives. In evaluations on the OPI benchmark, IBM reported that the LLMON preview models avoided 47% more attacks than the strongest tested frontier model and 95% more than the best tested open-source model, measured by the share of attacks the models did not follow. Similar trends were reported on ASB and StruQ. The company also stated that general performance remains on par with Granite 4.1 Instruct models across standard benchmark categories, suggesting that the security gains do not require a loss in broad capability. The models were trained from Granite 4.1 base checkpoints rather than existing instruct models, so LLMON behavior is learned during post-training. IBM re-serialized instruction-tuning data so each prompt component became an addressable block, and each turn included an execution block identifying the instruction, inputs, requirements, and output destination. Training also used decoy blocks left unbound to teach the model to follow declared dependencies while ignoring unrelated context. Later stages included GRPO across reasoning, RLHF, knowledge calibration, and mathematics, followed by preference tuning on responses to injection attempts. The resulting privilege rule treats instructions named in the exec block as authoritative, while data or tool outputs are treated as material to read rather than commands to follow. IBM emphasized that the separation is a learned behavior, not an enforced guarantee, and that work continues on adaptive attacks spanning multiple turns. LLMON is not limited to a fixed schema; developers can define domain-specific tags and fine-tune adapters. The release includes a ScienceWorld example in which a LoRA adapter learned a custom vocabulary for elementary science tasks, as well as notebooks for tool-calling agents and an attention-visualization tool. IBM described the models and tools as experimental previews for evaluation and feedback, not production deployment.

Related Links