来源时间 本站刊期
LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B
IBM's Granite 4.1 3B and 8B models add LLMON, an LLM-native markup language, which matters to developers building prompts and agent workflows who want structured output without brittle formatting hacks.
判断
AI platform teams already standardized on JSON should decide whether to pilot Granite 4.1 3B/8B LLMON previews for prompt-injection-sensitive agent workflows, accepting experimental preview status and conversion/tooling overhead before swapping production models.
依据
The release post says LLMON marks semantically coherent components with meaningful tags inside reserved special-token delimiters, uses unique IDs to bind instructions to data payloads, and ships JSON↔LLMON conversion tools. The two preview models are derived from Granite 4.1 3B/8B and use their own chat template, where a turn can mark parts exec: True or False; True parts are bound into an exec block. On OPI, the post reports LLMON avoids 47% more attacks than the best frontier model tested and 95% more than the best open-source model, while general benchmark categories are on par with Granite 4.1 Instruct.
- vs best frontier47% more attacks
- vs best open-source95% more
没写清 The source does not report absolute OPI robustness scores, latency or token costs beyond two-token block IDs, or production-support commitments for the preview models.
下一步 Check the GitHub toolkit's JSON↔LLMON round-trip examples and run the preview chat template on a held-out indirect-prompt-injection set before any migration.