In this paper, we propose a novel probability learning approach based on cross entropy (CE) to design the hybrid precoding with low-resolution phase shifters (PS) for URLLC users in a cell-free massive MIMO system. Based on the finite blocklength transmission, we construct a fairness optimization problem with the objective of maximizing the worst user’s achievable rate and minimizing the worst user’s decoding error probability successively. This kind of mixed integer programming problem is difficult to solve with traditional convex optimization, and the processing time is too long to meet the low delay requirements of URLLC due to the high complexity in cell-free systems. However, the CE-based approach we proposed can synchronously design the precoding of all distributed RRUs, adopting a parallel computing framework, and it greatly reduces processing latency, which is very friendly to URLLC service.