CO2 flooding is crucial for enhancing the oil recovery rate for low-permeability reservoirs. Optimizing injection and production rates offers advantages like low costs, easy implementation, and significant results. Current rates optimization methods can be categorized by objectives into single - objective and multiobjective, and by solution approaches into non - gradient and gradient methods. Non - gradient methods further split into classical machine learning and modern intelligent optimization techniques. These optimization methods comparison and applicable scenarios were investigated. Presently, rates optimization using deep learning - based reservoir proxy models is a new development trend, for they have computational efficiency thousands of times higher than traditional reservoir simulators, substantially cutting the time for numerous simulations in rates optimization. Yet, reservoir proxy models still face challenges, for the accuracy and reliability of proxy models require further enhancement. What’s more, Effective optimization algorithms are needed to tackle the complexity and uncertainty of reservoir systems. The integration physics-constrained proxy models and large language model presents a promising direction for rates optimization. The physics-constrained reservoir proxy models embedding mass conservation and Darcy’s law into the loss function to enable rapid and reliable prediction. A multi-objective optimization framework that couples reinforcement learning for real-time rates action generation with evolutionary algorithms for Pareto-front exploration, simultaneously considering oil recovery, carbon-storage efficiency and economic performance. Closed-loop, real-time optimization is thereby accomplished. The proposed model will markedly improves both computational efficiency and optimization performance, offering a practical scheme for the intelligent rates optimization of CO2 flooding.