Turbulent flow simulations pose significant challenges in computational fluid dynamics due to the huge variability in terms of spatial and temporal scales of the phenomenon. Direct Numerical Simulations (DNS), i.e., flow simulations on very fine meshes, provide a fully resolved representation of flow. However, they impose an unmanageable computational burden for most practical applications. To tackle this issue, under-resolved simulations on coarser meshes are often employed. To mitigate spurious oscillations that arise from the unresolved small-scale structures, the flow usually requires numerical regularization techniques to stabilize the simulation. In this talk, we focus on filter-based methods, such as the Evolve-Filter (EF) strategy, which have gained popularity due to their simplicity, modularity, and effectiveness. These approaches apply a differential filter with a proper filter radius to smooth the high-frequency components of the solution. However, when a large filter radius is used, EF may become over-diffusive, significantly compromising the solution accuracy. Selecting an appropriate filter parameter typically requires extensive, problem-specific tuning. We propose a novel approach to learn an optimal filter regularization strategy using reinforcement learning (RL). By framing the filtered simulation problem as a decision-making problem, an RL agent is trained to adjust the filter radius dynamically based on the evolving flow features. This allows an EF regularization with an improved accuracy solution and robust predictive capabilities.