Information reconciliation (IR) is an indispensable component in the post-processing stage of continuous-variable quantum key distribution (CV-QKD), which adopts error-correcting codes to address the asymmetry of secret keys. Currently, low-density parity-check (LDPC) decoding in IR is a post-processing bottleneck in high-speed CV-QKD systems since the upper bound on secret key rate is higher than the information throughput delivered by decoder. In this paper, we study the relationship between the syndrome variation pattern (SVP) in iterative decoding and reconciliation frame error rate. An early termination scheme based on SVP is proposed and applied to multidimensional reconciliation, which can increase information throughput by adaptively adjusting the iteration number of iterative decoding to real-time decoding status. Furthermore, we show that only the resulting syndrome of the highest-rate code part in Raptor-like LDPC codes needs to be calculated to verify whether the reconciliation is successful by studying the convergency of resulting syndrome, which can save a large fraction of computational resources for syndrome calculation. Simulation results show that information throughput of the proposed scheme can be improved by 617.1% compared to the existing scheme when the IR efficiency reaches 97.09%. The proposed scheme points out a new direction for breaking the post-processing bottleneck in high-speed CV-QKD systems.