The advent of the Internet of Things (IoT) enables billions of devices in wide-ranging domains such as healthcare, industry, and smart cities to interconnect with each other, but these connections make the network vulnerable to advanced cyber threats too. Current intrusion detection methods have failed to provide effective detection capabilities mainly because of issues such as extremely imbalanced data distributions, low classification accuracy, or static and manually tuned hyperparameters that do not generalize well in dynamic IoT settings. These challenges are exacerbated by unique IoT constraints, including limited device resources and dynamic attack patterns, which further complicate effective detection. To address these challenges, in this study we present a Reinforcement-enhanced Generative Artificial Intelligence (ReGAIN) framework for intelligent intrusion detection in IoT networks. In this approach, we use a generative autoencoder for data balancing to generate realistic minority class instances in the latent feature space, and meanwhile to obtain stable and unbiased learning of the model. This paper introduces a novel Pointer-Attention Dual Network (PAD-Net) that employs a Dual Attention Network (DANet) and a Pointer Network (PtrNet) to enhance spatial attention and inter-feature relationships. We also propose Reinforcement-enhanced PAD-Net (RePAD-Net), which leverages reinforcement learning to automatically optimize key hyperparameters at each training step, further enhancing generalization ability and robustness. The intrusion detection task in this study is a multi-class classification problem, where different types of attacks are distinguished from each other. Experimental results demonstrate that PAD-Net and RePAD-Net achieve notable improvements of 3.79% and 8.79% in accuracy, 3.79% and 8.78% in recall, 2.79% and 9.01% in F1-score, 3.79% and 8.83% in Mathews correlation coefficient, and 3.94% and 9.11% in Cohen’s Kappa, respectively, along with significant reductions in log loss of 47.42% and 70.96% and hamming loss of 24.33% and 56.37% compared with baseline models such as naive bayes, gradient boosting, densely connected network, long short term memory, hybrid models, DANet and PtrNet. Additionally, 10-fold cross validation is applied to validate the results of proposed models. These findings confirm that our proposed ReGAIN framework, which is able to alleviate data imbalance and improve learning generalization, can dramatically enhance the reliability of detection performance under complex IoT intrusion environments.