The technology relies on an XML-like markup that tags components such as requirements and data payloads with unique identifiers across different conversational roles. Two experimental models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, have been released alongside an open-source GitHub toolkit to demonstrate this capability. These models match the general performance of Granite 4.1 Instruct while reducing indirect prompt injection success rates to nearly zero on selected benchmarks. Integration is streamlined through lossless bi-directional JSON translation tools and native support within the Mellea library.
Key points
Using reserved special-token delimiters allows models to cleanly isolate untrusted data p…: Using reserved special-token delimiters allows models to cleanly isolate untrusted data payloads from binding instructions.
The markup works on top of existing chat templates rather than replacing them, enriching…: The markup works on top of existing chat templates rather than replacing them, enriching context across multiple conversational turns.
Native formatting support is implemented through the Mellea tool to facilitate adoption i…: Native formatting support is implemented through the Mellea tool to facilitate adoption in modern automated workflows.
Sources checked by our analysis pipeline: 1
LLMs rely on their context to generate the next set of tokens. Over a long trajectory, the context fills with numerous types of components: a system prompt, past conversation turns, instructions that act on retrieved documents, tool responses or user data, requirements constraining the allowed output, and reasoning chains. The context carries a rich structure, most of which is either implicit to the model or encoded in formats such as JSON that were not designed to be digested by LLMs. This is one of the main reasons LLMs, particularly those on the smaller end of the scale, are brittle and sometimes fail in unexpected ways. The lack of explicit structure can also be exploited by attackers, for example, in the form of prompt injections.
To address this, we developed LLMON (LLM Object Notation, pronounced “lemon”), an LLM-native markup language for explicitly communicating structure and semantics to the models. Semantically coherent components are marked with meaningful tags which are wrapped in reserved special-token delimiters. Data schemas, including nesting, are stated in a way that is unambiguous to a causal LLM. Unique…
Technical traces found within publicly accessible code indicate that multiple prominent Chinese artificial intelligence developers may be utilizing restricted hardware. Researchers uncovered evidence tying recently deployed systems to advanced computing components subject to international trade limits. This discovery raises fresh scrutiny regarding how restricted processing power continues to reach prominent foreign AI research labs.
Federal regulators are scaling automated advisory systems nationwide to address widespread flight bottlenecks and improve overall network flow. The technology rollout has earned strong endorsements from commercial carriers seeking relief from operational delays. However, the absence of workforce input introduces friction regarding how these recommendations will be integrated into day-to-day operations.
Meta is positioning its consumer AI agent to handle tedious real-world chores like waiting on hold with the DMV or booking restaurant reservations. The commercial push reflects the company's broader ambition to turn automated assistants into indispensable, paid productivity tools for everyday users. However, early friction remains high as businesses frequently terminate calls upon discovering an automated bot is on the line.