Abstract <p>Due to limited observational data and the inherent complexity of hydrological and climatic processes, accurately predicting runoff in data-scarce catchments-where only precipitation and potential evapotranspiration data are available-remains a major challenge. In such contexts, conceptual rainfall-runoff models like GR6J play a key role, as they are generally less sensitive to data length than data-driven models and capture the main hydrological processes governing streamflow generation (e.g., infiltration, soil moisture dynamics, groundwater exchange) through a parsimonious structure. This study investigates the effectiveness of the recently developed Kolmogorov-Arnold Network (KAN) in enhancing conceptual rainfall-runoff modeling under data-scarce conditions. A two-stage error-correction approach is proposed, in which the residual errors of the conceptual GR6J model are subsequently modeled using the KAN architecture. Furthermore, the effect of wavelet transform preprocessing is evaluated by decomposing the input data into multiple frequency components before feeding them into the KAN. Compared to the established LSTM-based models, the results show that KAN architectures underperform in standalone frameworks but generally outperform them in hybrid setups when applied to the Ouémé at Savè river basin in Benin and the Yala basin in Kenya. The GR6J-Wavelet-KAN (GR6J-WKAN) model achieves an NSE of 0.93, compared to 0.86 for GR6J alone in the Savè basin, and an NSE of 0.76 compared to 0.68 in the Yala basin. The GR6J-WKAN, followed by GR6J-WLSTM, accurately predicts total monthly runoff. Moreover, KAN-based hybrids show stronger performance than their corresponding LSTM-based ones for high flows. These findings underscore the potential of Kolmogorov-Arnold Networks-particularly when combined with appropriate wavelet-based preprocessing-to enhance predictive performance in conceptual rainfall-runoff modeling for data-scarce catchments.</p> Graphical Abstract <p>This study proposes hybrid models that couple the recently developed Kolmogorov-Arnold Network (KAN) and its wavelet variant (WKAN), as well as LSTM and WLSTM, with the widely used Génie Rural à 6 paramètres Journalier (GR6J) conceptual model to improve rainfall-runoff simulations in data-scarce catchments. The main objective is to assess the potential of KAN architectures for enhancing simulation performance under data-limited conditions. The graphical abstract provides a visual roadmap of the research workflow, starting from the meteorological forcing data used as inputs. These data include daily precipitation and evapotranspiration for training and testing the developed models. First, these inputs are used to generate daily runoff using the GR6J model. As shown in the graphical abstract, the residuals from GR6J are then modeled using KAN, LSTM, WKAN, and WLSTM architectures and used to error-correct the GR6J outputs, producing GR6J-KAN, GR6J-LSTM, GR6J-WKAN, and GR6J-WLSTM predictions. Several comparative analyses were conducted to assess the effectiveness of the KAN architectures relative to LSTM models. Overall performance was evaluated using performance metrics such as the Nash-Sutcliffe Efficiency (NSE), the Root Mean Square Error (RMSE), and Taylor diagram. In addition, total monthly runoffs were analyzed for accuracy, and extreme flow simulation accuracy was also assessed. The findings show the general superiority of GR6J-WKAN, although it is not always the best across all simulation scenarios-including low and high flows, total monthly runoff, and different basins-demonstrating that KAN architectures can substantially enhance water management in data-scarce catchments through hybrid modeling frameworks.</p>