The cutting-edge identification of scientific and technological literature is a key task in promoting scientific research innovation and technological development. However, traditional methods have two limitations: one is that they rely on both citation frequency as cutting-edge identifiers, resulting in insufficient robustness; the other is that they only use statistical data (such as citation networks) or text data (such as keywords and topics), and fail to fully integrate multi-source information, which limits the improvement of discrimination accuracy. In response to these issues, this study proposes an innovative cutting-edge identification framework. First, we designed a new cutting-edge index that breaks through the traditional single evaluation criteria for cited frequency and significantly improves the robustness of the identification. Secondly, we introduce the large language model (LLM) and propose a cutting-edge recognition method based on multimodal alignment. By deeply integrating text data (title, abstract) with statistical data (meta-information), we achieve more accurate discrimination. Experimental results show that the accuracy of this framework is more than 20% higher than that of traditional methods. This study not only provides an efficient and reliable tool for cutting-edge identification of scientific and technological literature, but also provides new ideas for the application of multimodal data fusion in scientific research and analysis, with important academic value and application prospects.