Preprint demonstrates AI agent optimizing wireless algorithms on baseband silicon

Researchers report in a preprint that an AI coding agent implemented and optimized wireless signal-processing algorithms on a commercial baseband chip, producing twelve kernels validated on silicon. The optimized versions took 18k-405k cycles (30-660 microseconds) per frame on the open 6G4All platform. The workflow guides the agent with measured hardware facts and checks its work against evaluators it is instructed never to edit. The agent also produced a detector requiring 13.9% fewer cycles than the port of an AI-discovered detector while matching its error rate. The researchers say their evaluation cannot statistically resolve the measured error-rate edge.