A new framework skips the network engineer and lets a language model translate business goals straight into working traffic rules.
Intent2Tc is a closed-loop system that takes a plain-language traffic-shaping request, breaks it into declarative sub-intents, and turns those into validated Linux tc (traffic control) configurations. The pipeline uses a digital-twin model built on Active Queue Management concepts to check that outputs behave as promised, plus a critique-and-refine loop and retrieval-augmented generation to reuse past translations. Researchers tested a mix of open-source large and small language models, alongside Claude Sonnet-4.6, against 100 traffic-shaping intents that comply with RFC 9315. Claude Sonnet-4.6 posted a 0.98 semantic similarity score, full coverage of required semantic units, and a normalized edit distance of just 0.045.
Intent-based networking has promised for years to let operators describe what they want instead of hand-writing configs, but the translation step from intent to executable rule has stayed stubbornly manual and error-prone. This paper is a data point that LLMs, paired with retrieval augmentation, can meaningfully close that gap - and that the RAG layer lets a much smaller model like Phi-4-mini approach the accuracy of the larger ones, which matters for anyone who cannot afford to run a frontier model on every config change.
Still, 100 test intents in a paper is a controlled demo, not a production network with legacy hardware, conflicting policies, and traffic that does not read the spec.