In recent years, deep learning has provided a new research path for block cipher security analysis. However, for large-state block ciphers, the high input dimensionality of neural distinguishers and the rapid expansion of the key search space severely limit the efficiency and stability of deep learning-assisted attacks. To address these issues, this study proposes DLAKRS, a deep learning-assisted key recovery scheme for large-state SIMECK block ciphers. In the training phase, the scheme introduces a gradient-based dynamic sensitive bit selection strategy to effectively compress the input dimension of the neural distinguisher while ensuring distinguishing performance. In the key recovery phase, a statistical score-guided search method is constructed based on the wrong-key response distribution to alleviate score fluctuations in the candidate key screening process. Based on the above scheme, this study constructs a multi-stage deep learning-assisted key recovery attack scheme for the SIMECK series of large-state lightweight block ciphers. The experimental results show that, when the input dimension is reduced by 45.8%–62.5%, the accuracy loss of the distinguisher is controlled within 10%, and the computational cost is significantly reduced compared with traditional multi-stage attack methods. In 16-round attacks on SIMECK48/96 and 18-round attacks on SIMECK64/128, the proposed scheme shows stable key recovery performance. These results show that the proposed scheme provides an effective way for deep learning-assisted key recovery of large-state SIMECK block ciphers.