AI Rewrites 5G Traffic Rules to Give You Faster Downloads
Your phone's data speeds could jump 44% — and engineers can now read exactly why.
When your phone feels slow, it's often because a cellular network is deciding who gets bandwidth and who waits. Engineers call this "network slicing" — think of it like dividing a highway into lanes, with some lanes reserved for emergency vehicles and hospitals that must never be blocked, and other lanes left open for regular traffic. Those decisions are usually made by an AI trained through trial and error, which works well but is essentially a black box: even the people running the network can't explain why it chose one lane over another.
A new paper from researchers Faezeh Dehghan Tarzjani and Bhaskar Krishnamachari at the University of Southern California takes a different route. They asked a large language model — the same kind of AI behind popular chatbots — to write the traffic-managing software as short, ordinary Python code instead of a tangled web of numbers. The AI proposes and revises versions, a simulator scores them, and the winner runs the network essentially unchanged. The result is a controller a human can actually read, like a recipe instead of a magic trick.
On the NSF POWDER 5G testbed, a real-world research network, the readable controller noticed when a guaranteed lane could no longer deliver its promised speed during a sustained signal fade. It released those wasted resources to ordinary users instead, lifting their speeds from 158.2 to 228.6 megabits per second — a 44.5% improvement over the best fixed setup. In one test, failures to meet service promises dropped from 79.9% to 2.2% after a single-line fix, because engineers could read the code and see the error.
The bigger idea: letting AI write code humans can inspect, rather than decisions humans can't. In a four-lane simulation, letting the AI evolve programs over many rounds beat one-shot requests by a wide margin — 51% average improvement when starting from an existing program, versus 32% from scratch and 16% from a single prompt. The tradeoff is that this only works where the network can be simulated accurately, and the approach still needs careful human review before it touches live networks.
- Instead of an unreadable AI black box, researchers had an AI write short, readable code to manage 5G traffic — so engineers can see and fix its decisions.
- In real 5G tests, regular download speeds rose 44.5%, from 158.2 to 228.6 Mbps, and failed service promises fell from 79.9% to 2.2%.
- Giving the AI many rounds to improve its code worked far better than asking once: 51% gains versus 16%.
Why It Matters
Faster, more reliable phone data — and network operators who can actually explain and fix outages instead of guessing.