There is no single best way to clean up a noisy 5G signal - so researchers built a dispatcher that picks the right tool for the moment.
A team tested eight existing channel-estimation methods, from simple least-squares interpolation to LMMSE and deep convolutional denoisers, across 3GPP-standard Urban-Macro, Urban-Micro, and Rural-Macro channel models built in NVIDIA's Sionna simulator. They ran the tests across both 5G-NR and LTE numerologies, and in single-antenna and 8x2 MIMO configurations. No single estimator won across the board; the best choice shifted with signal-to-noise ratio, mobility, and antenna setup. So the team built a condition-adaptive multi-agent orchestrator that treats each estimator as an independent agent and, using only a validation split rather than perfect foreknowledge, dispatches to whichever one is likely to perform best under the current conditions.
The orchestrator's picks landed within 1.07 dB of a theoretical oracle that always knows the right estimator in advance, and beat the best single fixed estimator by up to 3.6 dB at high SNR, exactly where the low-SNR favorite stops being useful. Because the eight estimators run independently, they can execute in parallel rather than in sequence: the study reports wall-clock time scaling down nearly linearly with worker count, up to a 6.9x speedup, which matters if this is ever going to run in a live base station rather than a research paper.
It is not a new estimator - it is a scheduling trick, and scheduling tricks are exactly the unglamorous engineering that turns a shelf of published methods into something a carrier might actually deploy.