[Purposes] In the field of bridge health monitoring, the precise identification of moving loads is crucial for bridge safety assessment. Existing methods still need improvement in terms of the modeling of temporal correlations in stochastic moving loads and handling the ill-posedness of inverse problems [Methods] This paper proposes a multi-point deflection response-driven first-order autoregressive (AR(1))-Tikhonov regularization-singular value decomposition (SVD)-generalized cross-validation (GCV) framework. First, temporally correlated stochastic loads factor sequences are generated via the AR(1) model and coupled with a baseline load to construct the stochastic moving loads model. Second, the ill-posed system is addressed by integrating Tikhonov regularization with SVD, and the regularization parameters are adaptively optimized using the GCV criterion. Finally, validation was performed using a simply supported beam finite element model with multi-point deflection data. [Findings] The results demonstrate that the relative percentage error (RPE) was 8.41% under noise-free conditions, which increased to 15.15% at a 2% noise level. Furthermore, increasing the number of measurement points improved the stochastic loads reconstruction accuracy, verifying the robustness of the framework. [Conclusions] This method breaks through the traditional deterministic identification framework and provides a new perspective for probabilistic bridge safety assessment.