Abstract: To evaluate the runoff simulation performance of traditional hydrological and machine learning models under identical observational data conditions in mountainous basins, this study takes the Qingjiang River Basin as a case study and compares the performance of the conceptual hydrological model (GR4J) and the Long Short-Term Memory network (LSTM) using the daily precipitation and runoff data from 2014 to 2023. Based on this, two hybrid models integrating physical mechanisms and data-driven approaches are developed, including a loosely coupled model (Hybrid1) and a tightly coupled model (Hybrid2). The improvements in simulation performance achieved by hybrid modeling strategies are further assessed. Results show that, under limited data conditions, when models are calibrated using single-year data, LSTM generally outperforms GR4J. When a combination of wet and dry years (2017 and 2019) is used for calibration, GR4J and LSTM achieve Nash-Sutcliffe efficiency (NSE) values of 0.742 and 0.790, and Kling-Gupta efficiency (KGE) values of 0.543 and 0.771 during the validation period, respectively. These results are significantly better than those obtained using single-year data and are comparable to those achieved with longer time series data (2014-2019). The hybrid modeling strategy further improves runoff simulation accuracy, with Hybrid2 showing better overall performance than Hybrid1. In Hybrid1, the runoff simulated by GR4J is used as an input feature for LSTM, representing a one-way information fusion strategy. By contrast, Hybrid2 not only outputs runoff through LSTM but also generates GR4J model parameters, which are subsequently fed back into GR4J, thereby forming a closed-loop interaction structure of "LSTM-parameters-GR4J". In particular, Hybrid2 attains NSE and KGE values of 0.795 and 0.771, respectively, during the validation period (2020-2023), representing improvements of 10.7% and 52.7% over GR4J, and 3.7% and 14.7% over LSTM. In addition, model performance exhibits clear seasonal variability, with all models achieving optimal accuracy in summer and the flood season. Hybrid2 achieves NSE and KGE values exceeding 0.8 in the flood season, demonstrating strong capability for flood-period runoff simulation. Overall, the results demonstrate the feasibility of training models with single-year and wet-dry year combination data, and confirm that the tightly coupled hybrid model effectively enhances runoff simulation accuracy in mountainous basins, providing valuable references for runoff simulation and flood forecasting.