<p>Estimating parameters for distributed hydrological and land-surface models is challenging, particularly in regions with limited observational data. One possible approach uses transfer functions that relate catchment attributes to model parameters, but these functions have so far been largely specified by hand, limiting flexibility and their practical use. Here we show that variational autoencoders can be used as text-generating models to automatically derive interpretable parameter transfer functions. This approach reformulates equation discovery as an optimization problem in a continuous latent space, improving both efficiency and transparency. We evaluate the method in a prediction-in-ungauged-basins setting using the mesoscale Hydrological Model across 162 German basins. The resulting transfer functions lead to improved runoff predictions compared with established regionalization methods and regional long short-term memory networks. In addition, the learned functions are robust across catchments, scalable to large spatial domains and maintain physical interpretability. These results demonstrate a pathway towards more transparent and transferable parameter estimation for large-scale process-based environmental models.</p>