[EMNLP 2025] Expanding before Inferring: Enhancing Factuality in Large Language Models through Premature Layers Interpolation