This study introduces an innovative EEG signal reconstruction sub-module designed to enhance the performance of deep learning models in EEG eye-tracking tasks. The sub-module integrates with Encoder-Classifier-based models and operates within a multi-task learning framework, enabling end-to-end training. It is trained under unsupervised conditions, making it versatile for various tasks. The effectiveness of the sub-module is demonstrated by incorporating it into advanced deep-learning models, including Transformers and pre-trained Transformers. The results show a significant improvement in feature representation capabilities, with a Root Mean Squared Error (RMSE) of 54.1mm, outperforming existing methods. The success of this approach suggests that the sub-module enhances the feature extraction ability of the encoder, preserving the end-to-end training process while saving computational costs associated with pre-training. The unsupervised nature of the sub-module allows for broader applicability across diverse tasks. The study highlights the potential of this novel paradigm in improving the performance of deep learning models in EEG-related challenges.This study introduces an innovative EEG signal reconstruction sub-module designed to enhance the performance of deep learning models in EEG eye-tracking tasks. The sub-module integrates with Encoder-Classifier-based models and operates within a multi-task learning framework, enabling end-to-end training. It is trained under unsupervised conditions, making it versatile for various tasks. The effectiveness of the sub-module is demonstrated by incorporating it into advanced deep-learning models, including Transformers and pre-trained Transformers. The results show a significant improvement in feature representation capabilities, with a Root Mean Squared Error (RMSE) of 54.1mm, outperforming existing methods. The success of this approach suggests that the sub-module enhances the feature extraction ability of the encoder, preserving the end-to-end training process while saving computational costs associated with pre-training. The unsupervised nature of the sub-module allows for broader applicability across diverse tasks. The study highlights the potential of this novel paradigm in improving the performance of deep learning models in EEG-related challenges.