GumbelRIS: End-to-End CNN-Based 2-Bit Phase Optimization for Large-Scale RIS-Aided MIMO Systems

We present GumbelRIS, a lightweight convolutional policy network that outputs discrete 2-bit RIS phase optimization in large-scale MIMO systems. It employs Gumbel-Softmax relaxation with a custom capacity-maximizing loss for differentiable training and hard quantization for inference. GumbelRIS outperforms metaheuristics and learning-based baselines across varying RIS sizes and channel conditions. It demonstrates superior capacity (126.757 bps/Hz), fast inference (28 ms), and robustness in NLOS scenarios. The model generalizes well without retraining and is scalable for practical deployment, making it suitable for dynamic and hardware-constrained RIS-aided 5G/6G environments.