What is it about?
The long-distance backbone of the internet runs on optical equipment built by many different manufacturers. Their products follow a shared standard, but working with them still requires knowing detailed technical rulebooks — called YANG models — that spell out exactly which settings each device will accept and which values are legal. Building that expertise takes years, and it keeps most researchers and engineers away from this equipment. This work builds a chat assistant that sits between a person and the optical network. Rather than trying to memorize the rulebooks, the assistant looks them up as it works: 154 rulebook files are broken into small pieces that keep their original structure, and the assistant decides for itself when a question needs a lookup and when it does not. Three separate checks then compare the values it produces against what the equipment actually permits, so it cannot invent a setting that sounds plausible but would be rejected by a real device. The assistant was tested on a free software simulator anyone can download, and then run live on a commercial network assembled from seven vendors' equipment at a major optical networking conference. The same software drove both without any changes.
Featured Image
Photo by Kirill Sh on Unsplash
Why is it important?
Optical networks are becoming central to AI infrastructure, since the training and serving of large models depends on moving data between and inside data centers. That pulls a new community of researchers toward optical transport — people who understand systems and machine learning, but not vendor-specific device models and configuration protocols. Language models are an obvious way to bridge that gap, but they do not know these device models, and a confidently wrong configuration is worse than no answer at all. This work shows a practical way to fix that: ground every answer in the actual standardized schema, and check it at three separate points before it reaches the user. The whole evaluation runs on freely available open-source software, so others can reproduce it without buying optical hardware, and the live commercial demonstration shows the same system works on real equipment.
Perspectives
What surprised me most was that the three grounding paths did not simply back each other up — each one caught a different kind of mistake. The post-generation check turned out to matter most exactly where retrieval was useless: when a user paraphrases a technical value in ordinary words, the term never appears in the schema index at all. Running the system live at OFC 2026, in front of people who could walk over and verify the answer against the equipment, was the real test.
Linqi Xiao
University of Texas at Dallas
Read the Original
This page is a summary of: A YANG-Grounded LLM Agent Supporting Multi-Vendor OpenROADM Optical Transport Networks, August 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3789240.3822606.
You can read the full text:
Contributors
The following have contributed to this page
