Field-level inference offers an optimal approach to extract information from cosmic structure and to reconstruct the initial conditions of the Universe. I will review different methods of field-level inference, ranging from differentiable forward modeling to machine learning approaches, with a particular focus on improving constraints from BAO reconstruction. I will also discuss and interpret which parts of the cosmic web neural networks pay most attention to during field-level inference, and explore the robustness of cosmological N-body simulations at the field level.