Intent Interpretation at RIC Timescales: Jev Decision Models versus Large Language Models in 6G Open RAN

Intent Interpretation at RIC Timescales. Jev Decision Models versus Large Language Models in 6G Open RAN. By Delong Li and four others, from the University of Technology Sydney. Published on the second of October. Its code repository is worked out: about one hundred and forty source files, with a readme and a licence.

Intent-based Open RAN needs an interpreter: something that turns an intent into an A1 policy, inside the loop of the RAN intelligent controller, or RIC. Two kinds of interpreter are on offer. Decision models such as Jev return typed policy fields. Generative large language models produce the policy token by token, which takes longer. The study asks whether that extra delay actually costs anything, in control deadlines, in RIC capacity, or in radio performance. Its answer comes in two parts. Slow interpreters do miss deadlines, and some saturate. But at the base point studied, no radio penalty from slow interpreters was resolved.

The comparison sets Jev and two other decision models against LLMs on the RANIntent v1 benchmark, both in closed-loop ns-3 simulation and on a real A1 and E2 path. Timing comes first. Median interpretation takes from about zero point two nine seconds up to two point three five seconds, while the A1 and E2 transfer itself takes under twenty-five milliseconds. Against the one-second near-real-time budget, Jev meets the deadline on ninety-nine point eight percent of calls. Two hosted LLMs meet it on seventeen point nine percent, and on zero percent. Capacity tells a similar story: two interpreters saturate their queues at two intents per second.

The radio network is where the picture changes. Ideal enforcement moves the service-level agreement violation for the affected class by three point nine six percentage points, in the direction each intent requests, compared with no update at the base point. Against that, no hosted LLM showed a resolved increase over Jev at that point. And per-second direct control gave no resolved reduction in SLA violation over a numerical xApp.

These results carry specific bounds. The radio comparisons are made at one base point, and the word that keeps recurring is resolved. The study did not resolve a radio penalty from slow interpreters there, which is a statement about what was detected at that point rather than a guarantee for every operating condition. The deadline result is measured against a one-second budget, and the queue result at a load of two intents per second. The models compared are Jev, two other decision models, and the LLMs tested, two of them hosted.

What separates the two kinds of interpreter here is time, and time shows up in deadlines and queues before it shows up in the radio. Measured against the one-second budget, typed decisions hold and token-by-token generation mostly does not, while the radio network itself, at the point studied, did not register a resolved difference.