Fig. 1. A general quantum metrology device consists of quantum state preparation, parameter-encoding evolution and measurement. Here, the device is equipped with an ML agent. The agent associates to each measurement result μ a Bayesian distribution P(ϕ|μ) obtained from a neural network trained with calibration data. When acquiring repeated measurement with results μ1,…μm, Bayesian distributions are multiplied, updating the prior knowledge about the unknown parameter ϕT. Finally, the agent chooses computes a phase ϕC for an adaptive feedback control of the device. Notice that ϕT and ϕC can be single- (as in the schematic here) or multivalued (as in the experiment of Cimini et al.¹⁷).