Abstract: In mega-constellation laser inter-satellite link missions, uncertainty factors—such as orbital dynamics, attitude dynamics, space environmental disturbances, and laser terminal errors—are unavoidable. These factors often lead to initial acquisition difficulties and challenges in maintaining long-term link stability. Consequently, systematic identification and effective correction of dynamic pointing errors in inter-satellite links are urgently required. To address this issue, this paper proposes a lightweight, adaptive, and robust error parameter identification and correction framework based on Bayesian inference, featuring efficient forward simulation capabilities and adaptive precision sampling. First, based on on-orbit telemetry statistics and ground-based simulation results, the primary error sources are systematically analyzed and quantified. Second, a State-Space Gaussian Process Regression (SS-GPR) method is proposed to construct a surrogate model to characterize the error dynamics while a physics-informed spatiotemporal kernel is designed. Compared with the high-fidelity simulation model, this SS-GPR surrogate reduces the single prediction execution time from seconds to milliseconds, achieving an average relative error of less than 1% on the test set. Furthermore, an Adaptive Robust Tempered Markov Chain Monte Carlo (ART-MCMC) algorithm is developed, capable of achieving rapid convergence to complete the posterior inference of high-dimensional error parameters. Its average normalized calibration deviation is below 5%, which mitigates the limitations of conventional Markov Chain Monte Carlo methods in high-dimensional parameter spaces, such as low sampling efficiency, slow convergence, and tuning difficulties. Finally, the identified parameter calibration results are utilized to calibrate the dynamic model. Experimental data demonstrate that after correction, the average errors of elevation and azimuth angles are reduced from 0.00027° and 0.00148° to 0.000006° and 0.000034°, respectively, with the residual distribution converging to a zero-mean distribution with significantly reduced variance. Simulation results verify that the lightweight Bayesian inference-based identification and correction framework proposed in this paper effectively addresses the dynamic pointing errors in inter-satellite links induced by multi-source uncertainties.