This paper introduces JAG-R, a novel online geometric recentering method to enhance the performance of neural networks, specifically complex-valued neural networks (CVNNs), on noisy exclusive or (XOR) classification tasks. The proposed method optimises the placement of data points by maximising angular separability, thereby improving classification accuracy under various noise conditions. Experimental results demonstrate that the proposed approach significantly reduces classification errors for two types of noise: independent Gaussian noise, and morphing noise. Specifically, under independent Gaussian noise, the JAG-R method achieves a 22.27% improvement over the Centroid method and a 58.04% improvement over a Baseline at the largest bias variance. For morphing noise, also at the largest bias variance, JAG-R outperforms Centroid by 31.86% and the Baseline by 50.4%. These findings highlight the potential of the JAG-R method for improving complex-valued neural network performance in noisy environments, with implications for real-time signal processing, online learning and other applications involving structured data distortions.